chore: update lance dependency to v8.0.0-beta.9

## Summary

Fixes the `NAPI_RS_FORCE_WASI=false` issue by upgrading `@napi-rs/cli`
from `3.5.1` to `3.7.0`.

Closes #3267

## Root Cause

In the `native.js` loader generated by `napi build`, the check was:

```js
if (!nativeBinding || process.env.NAPI_RS_FORCE_WASI) {
```

In JavaScript, any non-empty string is truthy, so
`NAPI_RS_FORCE_WASI=false` (a non-empty string) inadvertently triggered
the WASI fallback path. This caused an `ENOENT` error when
`lancedb.wasi.cjs` was not present.

## Fix

`@napi-rs/cli@3.7.0`
([napi-rs/napi-rs#3236](https://github.com/napi-rs/napi-rs/pull/3236))
introduced a tri-state check in the template that generates `native.js`:

**Before (generated by @napi-rs/cli@3.5.1):**
```js
if (!nativeBinding || process.env.NAPI_RS_FORCE_WASI) {
```

**After (generated by @napi-rs/cli@3.7.0):**
```js
const forceWasi =
  process.env.NAPI_RS_FORCE_WASI === 'true' || process.env.NAPI_RS_FORCE_WASI === 'error'

if (!nativeBinding || forceWasi) {
```

Only the literal string `'true'` (or `'error'` for strict mode) now
activates the WASI path. All other values, including `'false'`, `'0'`,
or an unset variable, behave as if WASI is not forced.

## Changes

- `nodejs/package.json`: bump `@napi-rs/cli` from `3.5.1` to `3.7.0`
- `nodejs/package-lock.json` / `nodejs/pnpm-lock.yaml`: update lock
files to match

The fix is in the upstream napi-rs tool; the generated `native.js` is
not committed to this repository and is produced at build time by `napi
build`.

.agents/skills/README.md

@@ -0,0 +1,7 @@
+# Agent Skills
+
+This directory contains repo-scoped code agent skills for the LanceDB project.
+
+Each skill is a folder that contains a required `SKILL.md` and optional bundled resources.
+
+Codex discovers skills from `.agents/skills` in the current working directory and parent directories.

.agents/skills/lancedb-update-lance-dependency/SKILL.md

@@ -0,0 +1,98 @@
+---
+name: lancedb-update-lance-dependency
+description: Update LanceDB to a specific Lance release or tag. Use when bumping Lance dependencies in the lancedb repository, including Rust workspace Lance crates, Java lance-core, validation, branch creation, commit, push, and PR creation when requested.
+---
+
+# LanceDB Update Lance Dependency
+
+## Scope
+
+Use this skill in the `lancedb/lancedb` repository when updating the Lance dependency to a specific Lance version or tag.
+
+Inputs can be a version (`7.2.0-beta.1`), a tag (`v7.2.0-beta.1`), a tag ref (`refs/tags/v7.2.0-beta.1`), or `latest`.
+
+## Workflow
+
+1. Confirm the worktree status with `git status --short`.
+2. Resolve the target Lance version:
+
+   - If the input is `latest`, empty, or omitted, run:
+
+     ```bash
+     python3 ci/check_lance_release.py
+     ```
+
+     Parse the JSON output. If `needs_update` is not `true`, stop without creating a PR. Otherwise use `latest_tag`.
+
+   - If the input is explicit, use it directly.
+
+3. Compute update metadata without changing files:
+
+   ```bash
+   python3 ci/update_lance_dependency.py "$TAG_OR_VERSION" --metadata-only
+   ```
+
+   Before making changes, check for an existing open PR with the emitted `pr_title`:
+
+   ```bash
+   gh pr list --search "\"$PR_TITLE\" in:title" --state open --limit 1 --json number,url,title
+   ```
+
+   If a matching open PR exists, stop and report it instead of creating a duplicate.
+
+4. Run the deterministic update entrypoint:
+
+   ```bash
+   python3 ci/update_lance_dependency.py "$TAG_OR_VERSION"
+   ```
+
+   This updates the Rust workspace Lance dependencies through `ci/set_lance_version.py`, updates `java/pom.xml`, refreshes Cargo metadata, and prints JSON metadata containing `branch_name`, `commit_message`, and `pr_title`.
+
+5. Run validation:
+
+   ```bash
+   cargo clippy --quiet --workspace --tests --all-features -- -D warnings
+   cargo fmt --all --quiet
+   ```
+
+   Fix real diagnostics and rerun clippy until it succeeds. Do not skip warnings.
+
+6. Inspect `git status --short` and `git diff` to ensure only the Lance dependency update and required compatibility fixes are present.
+
+7. If the task only asks to prepare local changes, stop here and report the changed files and validation result.
+
+8. If the task asks to publish the update, create a branch using the printed `branch_name`, stage all relevant files, and commit using the printed `commit_message`. Do not amend or rewrite existing commits.
+
+9. Push to `origin`. Before creating the PR, check that the current token has push permission:
+
+   ```bash
+   gh api repos/lancedb/lancedb --jq .permissions.push
+   ```
+
+   If the remote branch already exists for the same generated branch name, delete the remote ref with `gh api -X DELETE repos/lancedb/lancedb/git/refs/heads/$BRANCH_NAME`, then push. Do not force-push.
+
+10. Create a PR targeting `main` with the printed `pr_title`. If there is no PR template, keep the body to two or three concise sentences: state the Lance dependency bump, note any required compatibility fixes, and link the triggering Lance tag or release.
+
+11. Read back the remote PR title after creation. If it is not a Conventional Commit title, fix it immediately.
+
+12. When running in GitHub Actions after creating the LanceDB PR, trigger the Sophon dependency update:
+
+    ```bash
+    gh workflow run codex-bump-lancedb-lance.yml \
+      --repo lancedb/sophon \
+      -f lance_ref="$LANCE_TAG" \
+      -f lancedb_ref="$BRANCH_NAME"
+    gh run list --repo lancedb/sophon --workflow codex-bump-lancedb-lance.yml --limit 1 --json databaseId,url,displayTitle
+    ```
+
+    Use the emitted metadata `tag` value as `LANCE_TAG`. Do this only after a new LanceDB PR has been created. If the update was skipped because no update is needed or an open PR already exists, do not trigger Sophon.
+
+## GitHub Actions
+
+When this skill is used from GitHub Actions, `TAG`, `GH_TOKEN`, and `GITHUB_TOKEN` may already be set. Resolve `latest` first when `TAG` is empty. Once an explicit tag or version is known, use:
+
+```bash
+python3 ci/update_lance_dependency.py "$TAG" --github-output "$GITHUB_OUTPUT"
+```
+
+Then use the emitted `branch_name`, `commit_message`, and `pr_title` values for branch, commit, and PR creation.

.bumpversion.toml

@@ -1,5 +1,5 @@
 [tool.bumpversion]
-current_version = "0.30.0-beta.0"
+current_version = "0.30.1-beta.2"
 parse = """(?x)
     (?P<major>0|[1-9]\\d*)\\.
     (?P<minor>0|[1-9]\\d*)\\.

.github/dependabot.yml

@@ -21,3 +21,14 @@ updates:
         update-types:
           - minor
           - patch
+
+  - package-ecosystem: pip
+    directory: /python
+    schedule:
+      interval: weekly
+    # Only update uv.lock, never widen version requirements in pyproject.toml.
+    versioning-strategy: lockfile-only
+    groups:
+      python-deps:
+        patterns:
+          - "*"

.github/workflows/build_windows_wheel/action.yml

@@ -29,7 +29,3 @@ runs:
         args: ${{ inputs.args }}
         docker-options: "-e PIP_EXTRA_INDEX_URL='https://pypi.fury.io/lance-format/ https://pypi.fury.io/lancedb/'"
         working-directory: python
-    - uses: actions/upload-artifact@v4
-      with:
-        name: windows-wheels
-        path: python\target\wheels

.github/workflows/codex-update-lance-dependency.yml

@@ -4,14 +4,16 @@ on:
   workflow_call:
     inputs:
       tag:
-        description: "Tag name from Lance"
-        required: true
+        description: "Tag name from Lance. If omitted, the skill will use the latest Lance release that needs an update."
+        required: false
+        default: ""
         type: string
   workflow_dispatch:
     inputs:
       tag:
-        description: "Tag name from Lance"
-        required: true
+        description: "Tag name from Lance. Leave empty to use the latest Lance release that needs an update."
+        required: false
+        default: ""
         type: string
 
 permissions:
@@ -25,7 +27,7 @@ jobs:
     steps:
       - name: Show inputs
         run: |
-          echo "tag = ${{ inputs.tag }}"
+          echo "tag = ${{ inputs.tag || 'latest' }}"
 
       - name: Checkout Repo LanceDB
         uses: actions/checkout@v4
@@ -71,65 +73,21 @@ jobs:
           OPENAI_API_KEY: ${{ secrets.CODEX_TOKEN }}
         run: |
           set -euo pipefail
-          VERSION="${TAG#refs/tags/}"
-          VERSION="${VERSION#v}"
-          BRANCH_NAME="codex/update-lance-${VERSION//[^a-zA-Z0-9]/-}"
-
-          # Use "chore" for beta/rc versions, "feat" for stable releases
-          if [[ "${VERSION}" == *beta* ]] || [[ "${VERSION}" == *rc* ]]; then
-            COMMIT_TYPE="chore"
-          else
-            COMMIT_TYPE="feat"
-          fi
+          TARGET_TAG="${TAG:-latest}"
 
           cat <<EOF >/tmp/codex-prompt.txt
-          You are running inside the lancedb repository on a GitHub Actions runner. Update the Lance dependency to version ${VERSION} and prepare a pull request for maintainers to review.
+          You are running inside the lancedb repository on a GitHub Actions runner.
 
-          Follow these steps exactly:
-          1. Use script "ci/set_lance_version.py" to update Lance Rust dependencies. The script already refreshes Cargo metadata, so allow it to finish even if it takes time.
-          2. Update the Java lance-core dependency version in "java/pom.xml": change the "<lance-core.version>...</lance-core.version>" property to "${VERSION}".
-          3. Run "cargo clippy --workspace --tests --all-features -- -D warnings". If diagnostics appear, fix them yourself and rerun clippy until it exits cleanly. Do not skip any warnings.
-          4. After clippy succeeds, run "cargo fmt --all" to format the workspace.
-          5. Ensure the repository is clean except for intentional changes. Inspect "git status --short" and "git diff" to confirm the dependency update and any required fixes.
-          6. Create and switch to a new branch named "${BRANCH_NAME}" (replace any duplicated hyphens if necessary).
-          7. Stage all relevant files with "git add -A". Commit using the message "${COMMIT_TYPE}: update lance dependency to v${VERSION}".
-          8. Push the branch to origin. If the remote branch already exists, delete it first with "gh api -X DELETE repos/lancedb/lancedb/git/refs/heads/${BRANCH_NAME}" then push with "git push origin ${BRANCH_NAME}". Do NOT use "git push --force" or "git push -f".
-          9. env "GH_TOKEN" is available, use "gh" tools for github related operations like creating pull request.
-          10. Create a pull request targeting "main" with title "${COMMIT_TYPE}: update lance dependency to v${VERSION}". First, write the PR body to /tmp/pr-body.md using a heredoc (cat <<'EOF' > /tmp/pr-body.md). The body should summarize the dependency bump, clippy/fmt verification, and link the triggering tag (${TAG}). Then run "gh pr create --body-file /tmp/pr-body.md".
-          11. After creating the PR, display the PR URL, "git status --short", and a concise summary of the commands run and their results.
+          Use \$lancedb-update-lance-dependency with target "${TARGET_TAG}".
 
           Constraints:
-          - Use bash commands; avoid modifying GitHub workflow files other than through the scripted task above.
-          - Do not merge the PR.
-          - If any command fails, diagnose and fix the issue instead of aborting.
+          - Use env "GH_TOKEN" for GitHub operations.
+          - Do not merge the pull request.
+          - Do not force-push.
+          - Do not create a duplicate pull request if an open PR already exists for the target Lance version.
+          - If any command fails, diagnose and fix the root cause instead of aborting.
+          - After creating the PR, display the PR URL, "git status --short", and a concise summary of the commands run and their results.
           EOF
 
           printenv OPENAI_API_KEY | codex login --with-api-key
           codex --config shell_environment_policy.ignore_default_excludes=true exec --dangerously-bypass-approvals-and-sandbox "$(cat /tmp/codex-prompt.txt)"
-
-      - name: Trigger sophon dependency update
-        env:
-          TAG: ${{ inputs.tag }}
-          GH_TOKEN: ${{ secrets.ROBOT_TOKEN }}
-        run: |
-          set -euo pipefail
-          VERSION="${TAG#refs/tags/}"
-          VERSION="${VERSION#v}"
-          LANCEDB_BRANCH="codex/update-lance-${VERSION//[^a-zA-Z0-9]/-}"
-
-          echo "Triggering sophon workflow with:"
-          echo "  lance_ref: ${TAG#refs/tags/}"
-          echo "  lancedb_ref: ${LANCEDB_BRANCH}"
-
-          gh workflow run codex-bump-lancedb-lance.yml \
-            --repo lancedb/sophon \
-            -f lance_ref="${TAG#refs/tags/}" \
-            -f lancedb_ref="${LANCEDB_BRANCH}"
-
-      - name: Show latest sophon workflow run
-        env:
-          GH_TOKEN: ${{ secrets.ROBOT_TOKEN }}
-        run: |
-          set -euo pipefail
-          echo "Latest sophon workflow run:"
-          gh run list --repo lancedb/sophon --workflow codex-bump-lancedb-lance.yml --limit 1 --json databaseId,url,displayTitle

.github/workflows/lance-release-timer.yml

@@ -1,62 +0,0 @@
-name: Lance Release Timer
-
-on:
-  schedule:
-    - cron: "*/10 * * * *"
-  workflow_dispatch:
-
-permissions:
-  contents: read
-  actions: write
-
-concurrency:
-  group: lance-release-timer
-  cancel-in-progress: false
-
-jobs:
-  trigger-update:
-    runs-on: ubuntu-latest
-    steps:
-      - name: Checkout repository
-        uses: actions/checkout@v4
-
-      - name: Check for new Lance tag
-        id: check
-        env:
-          GH_TOKEN: ${{ secrets.ROBOT_TOKEN }}
-        run: |
-          python3 ci/check_lance_release.py --github-output "$GITHUB_OUTPUT"
-
-      - name: Look for existing PR
-        if: steps.check.outputs.needs_update == 'true'
-        id: pr
-        env:
-          GH_TOKEN: ${{ secrets.ROBOT_TOKEN }}
-        run: |
-          set -euo pipefail
-          TITLE="chore: update lance dependency to v${{ steps.check.outputs.latest_version }}"
-          COUNT=$(gh pr list --search "\"$TITLE\" in:title" --state open --limit 1 --json number --jq 'length')
-          if [ "$COUNT" -gt 0 ]; then
-            echo "Open PR already exists for $TITLE"
-            echo "pr_exists=true" >> "$GITHUB_OUTPUT"
-          else
-            echo "No existing PR for $TITLE"
-            echo "pr_exists=false" >> "$GITHUB_OUTPUT"
-          fi
-
-      - name: Trigger codex update workflow
-        if: steps.check.outputs.needs_update == 'true' && steps.pr.outputs.pr_exists != 'true'
-        env:
-          GH_TOKEN: ${{ secrets.ROBOT_TOKEN }}
-        run: |
-          set -euo pipefail
-          TAG=${{ steps.check.outputs.latest_tag }}
-          gh workflow run codex-update-lance-dependency.yml -f tag=refs/tags/$TAG
-
-      - name: Show latest codex workflow run
-        if: steps.check.outputs.needs_update == 'true' && steps.pr.outputs.pr_exists != 'true'
-        env:
-          GH_TOKEN: ${{ secrets.ROBOT_TOKEN }}
-        run: |
-          set -euo pipefail
-          gh run list --workflow codex-update-lance-dependency.yml --limit 1 --json databaseId,url,displayTitle

.github/workflows/pypi-publish.yml

@@ -8,6 +8,9 @@ on:
     # This should trigger a dry run (we skip the final publish step)
     paths:
       - .github/workflows/pypi-publish.yml
+      - .github/workflows/build_linux_wheel/action.yml
+      - .github/workflows/build_mac_wheel/action.yml
+      - .github/workflows/build_windows_wheel/action.yml
       - Cargo.toml # Change in dependency frequently breaks builds
       - Cargo.lock
 
@@ -21,32 +24,21 @@ jobs:
   linux:
     name: Python ${{ matrix.config.platform }} manylinux${{ matrix.config.manylinux }}
     timeout-minutes: 60
-    permissions:
-      id-token: write
-      contents: read
     strategy:
       matrix:
         config:
-          - platform: x86_64
-            manylinux: "2_17"
-            extra_args: ""
-            runner: ubuntu-22.04
           - platform: x86_64
             manylinux: "2_28"
             extra_args: "--features fp16kernels"
             runner: ubuntu-22.04
-          - platform: aarch64
-            manylinux: "2_17"
-            extra_args: ""
-            # For successful fat LTO builds, we need a large runner to avoid OOM errors.
-            runner: ubuntu-2404-8x-arm64
+          # For successful fat LTO builds, we need a large runner to avoid OOM errors.
           - platform: aarch64
             manylinux: "2_28"
             extra_args: "--features fp16kernels"
             runner: ubuntu-2404-8x-arm64
     runs-on: ${{ matrix.config.runner }}
     steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v6
         with:
           fetch-depth: 0
           lfs: true
@@ -60,15 +52,14 @@ jobs:
           args: "--release --strip ${{ matrix.config.extra_args }}"
           arm-build: ${{ matrix.config.platform == 'aarch64' }}
           manylinux: ${{ matrix.config.manylinux }}
-      - uses: ./.github/workflows/upload_wheel
+      - uses: actions/upload-artifact@v7
         if: startsWith(github.ref, 'refs/tags/python-v')
         with:
-          fury_token: ${{ secrets.FURY_TOKEN }}
+          name: wheels-linux-${{ matrix.config.platform }}-${{ matrix.config.manylinux }}
+          path: target/wheels/lancedb-*.whl
+          if-no-files-found: error
   mac:
     timeout-minutes: 90
-    permissions:
-      id-token: write
-      contents: read
     runs-on: ${{ matrix.config.runner }}
     strategy:
       matrix:
@@ -78,7 +69,7 @@ jobs:
     env:
       MACOSX_DEPLOYMENT_TARGET: 10.15
     steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v6
         with:
           fetch-depth: 0
           lfs: true
@@ -90,18 +81,21 @@ jobs:
         with:
           python-minor-version: 10
           args: "--release --strip --target ${{ matrix.config.target }} --features fp16kernels"
-      - uses: ./.github/workflows/upload_wheel
+      - uses: actions/upload-artifact@v7
         if: startsWith(github.ref, 'refs/tags/python-v')
         with:
-          fury_token: ${{ secrets.FURY_TOKEN }}
+          name: wheels-mac-${{ matrix.config.target }}
+          path: target/wheels/lancedb-*.whl
+          if-no-files-found: error
   windows:
-    timeout-minutes: 60
-    permissions:
-      id-token: write
-      contents: read
+    timeout-minutes: 90
     runs-on: windows-latest
+    env:
+      # link.exe is single-threaded and the long pole on Windows builds. Use
+      # rustc's bundled lld-link instead.
+      CARGO_TARGET_X86_64_PC_WINDOWS_MSVC_LINKER: rust-lld
     steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v6
         with:
           fetch-depth: 0
           lfs: true
@@ -113,18 +107,70 @@ jobs:
         with:
           python-minor-version: 10
           args: "--release --strip"
-          vcpkg_token: ${{ secrets.VCPKG_GITHUB_PACKAGES }}
-      - uses: ./.github/workflows/upload_wheel
+      - uses: actions/upload-artifact@v7
         if: startsWith(github.ref, 'refs/tags/python-v')
         with:
-          fury_token: ${{ secrets.FURY_TOKEN }}
+          name: wheels-windows
+          path: target/wheels/lancedb-*.whl
+          if-no-files-found: error
+  publish:
+    name: Publish wheels
+    if: startsWith(github.ref, 'refs/tags/python-v')
+    needs: [linux, mac, windows]
+    runs-on: ubuntu-latest
+    permissions:
+      id-token: write
+      contents: read
+    steps:
+      - uses: actions/checkout@v6
+      - name: Download wheel artifacts
+        uses: actions/download-artifact@v8
+        with:
+          pattern: wheels-*
+          path: target/wheels
+          merge-multiple: true
+      - name: List wheels
+        run: ls -la target/wheels
+      - name: Choose repo
+        id: choose_repo
+        run: |
+          if [[ ${{ github.ref }} == *beta* ]]; then
+            echo "repo=fury" >> $GITHUB_OUTPUT
+          else
+            echo "repo=pypi" >> $GITHUB_OUTPUT
+          fi
+      - name: Publish to Fury
+        if: steps.choose_repo.outputs.repo == 'fury'
+        env:
+          FURY_TOKEN: ${{ secrets.FURY_TOKEN }}
+        run: |
+          shopt -s nullglob
+          WHEELS=(target/wheels/lancedb-*.whl)
+          if [[ ${#WHEELS[@]} -eq 0 ]]; then
+            echo "No wheels found in target/wheels/" >&2
+            exit 1
+          fi
+          for WHEEL in "${WHEELS[@]}"; do
+            echo "Uploading $WHEEL to Fury"
+            curl -f -F package=@"$WHEEL" "https://$FURY_TOKEN@push.fury.io/lancedb/"
+          done
+      # NOTE: pypa/gh-action-pypi-publish must be invoked directly from a
+      # workflow file, not from inside a composite action. When called from a
+      # composite, `github.action_repository` is empty (actions/runner#2473)
+      # and the action falls back to `github.repository`, producing a bogus
+      # `docker://ghcr.io/<repo>:<ref>` image reference that GHA tries to pull.
+      - name: Publish to PyPI
+        if: steps.choose_repo.outputs.repo == 'pypi'
+        uses: pypa/gh-action-pypi-publish@release/v1
+        with:
+          packages-dir: target/wheels/
   gh-release:
     if: startsWith(github.ref, 'refs/tags/python-v')
     runs-on: ubuntu-latest
     permissions:
       contents: write
     steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v6
         with:
           fetch-depth: 0
           lfs: true
@@ -187,13 +233,13 @@ jobs:
   report-failure:
     name: Report Workflow Failure
     runs-on: ubuntu-latest
-    needs: [linux, mac, windows]
+    needs: [linux, mac, windows, publish]
     permissions:
       contents: read
       issues: write
     if: always() && failure() && startsWith(github.ref, 'refs/tags/python-v')
     steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v6
       - uses: ./.github/actions/create-failure-issue
         with:
           job-results: ${{ toJSON(needs) }}

.github/workflows/upload_wheel/action.yml

@@ -1,34 +0,0 @@
-name: upload-wheel
-
-description: "Upload wheels to Pypi"
-inputs:
-  fury_token:
-    required: true
-    description: "release token for the fury repo"
-
-runs:
-  using: "composite"
-  steps:
-  - name: Choose repo
-    shell: bash
-    id: choose_repo
-    run: |
-      if [[ ${{ github.ref }} == *beta* ]]; then
-        echo "repo=fury" >> $GITHUB_OUTPUT
-      else
-        echo "repo=pypi" >> $GITHUB_OUTPUT
-      fi
-  - name: Publish to Fury
-    if: steps.choose_repo.outputs.repo == 'fury'
-    shell: bash
-    env:
-      FURY_TOKEN: ${{ inputs.fury_token }}
-    run: |
-      WHEEL=$(ls target/wheels/lancedb-*.whl 2> /dev/null | head -n 1)
-      echo "Uploading $WHEEL to Fury"
-      curl -f -F package=@$WHEEL https://$FURY_TOKEN@push.fury.io/lancedb/
-  - name: Publish to PyPI
-    if: steps.choose_repo.outputs.repo == 'pypi'
-    uses: pypa/gh-action-pypi-publish@release/v1
-    with:
-      packages-dir: target/wheels/

AGENTS.md

@@ -37,10 +37,13 @@ Before committing changes, run formatting for every language you touched. At min
   and run targeted tests through `cd python && uv run ...`.
 * TypeScript changes: run the relevant `npm`/`pnpm` lint, format, build, and docs commands in `nodejs`.
 
-Before creating a PR, make sure the PR title follows Conventional Commits, such as
-`fix: support nested field paths in native index creation` or
-`feat(python): add dataset multiprocessing support`. The semantic-release check uses the
-PR title and body as the merge commit message, so a non-conventional PR title will fail CI.
+Before creating a PR, the exact value passed to `gh pr create --title` must follow
+Conventional Commits, such as `fix: support nested field paths in native index creation`
+or `feat(python): add dataset multiprocessing support`. Do not use a plain natural
+language summary like `Support nested field paths in native index creation` as the PR
+title. The semantic-release check uses the PR title and body as the merge commit message,
+so a non-conventional PR title will fail CI. After creating a PR, read the remote PR title
+back and fix it immediately if it is not conventional.
 
 ## Coding tips
 

Cargo.toml

@@ -13,20 +13,20 @@ categories = ["database-implementations"]
 rust-version = "1.91.0"
 
 [workspace.dependencies]
-lance = { "version" = "=7.0.0-beta.13", default-features = false, "tag" = "v7.0.0-beta.13", "git" = "https://github.com/lance-format/lance.git" }
-lance-core = { "version" = "=7.0.0-beta.13", "tag" = "v7.0.0-beta.13", "git" = "https://github.com/lance-format/lance.git" }
-lance-datagen = { "version" = "=7.0.0-beta.13", "tag" = "v7.0.0-beta.13", "git" = "https://github.com/lance-format/lance.git" }
-lance-file = { "version" = "=7.0.0-beta.13", "tag" = "v7.0.0-beta.13", "git" = "https://github.com/lance-format/lance.git" }
-lance-io = { "version" = "=7.0.0-beta.13", default-features = false, "tag" = "v7.0.0-beta.13", "git" = "https://github.com/lance-format/lance.git" }
-lance-index = { "version" = "=7.0.0-beta.13", "tag" = "v7.0.0-beta.13", "git" = "https://github.com/lance-format/lance.git" }
-lance-linalg = { "version" = "=7.0.0-beta.13", "tag" = "v7.0.0-beta.13", "git" = "https://github.com/lance-format/lance.git" }
-lance-namespace = { "version" = "=7.0.0-beta.13", "tag" = "v7.0.0-beta.13", "git" = "https://github.com/lance-format/lance.git" }
-lance-namespace-impls = { "version" = "=7.0.0-beta.13", default-features = false, "tag" = "v7.0.0-beta.13", "git" = "https://github.com/lance-format/lance.git" }
-lance-table = { "version" = "=7.0.0-beta.13", "tag" = "v7.0.0-beta.13", "git" = "https://github.com/lance-format/lance.git" }
-lance-testing = { "version" = "=7.0.0-beta.13", "tag" = "v7.0.0-beta.13", "git" = "https://github.com/lance-format/lance.git" }
-lance-datafusion = { "version" = "=7.0.0-beta.13", "tag" = "v7.0.0-beta.13", "git" = "https://github.com/lance-format/lance.git" }
-lance-encoding = { "version" = "=7.0.0-beta.13", "tag" = "v7.0.0-beta.13", "git" = "https://github.com/lance-format/lance.git" }
-lance-arrow = { "version" = "=7.0.0-beta.13", "tag" = "v7.0.0-beta.13", "git" = "https://github.com/lance-format/lance.git" }
+lance = { "version" = "=8.0.0-beta.9", default-features = false, "tag" = "v8.0.0-beta.9", "git" = "https://github.com/lance-format/lance.git" }
+lance-core = { "version" = "=8.0.0-beta.9", "tag" = "v8.0.0-beta.9", "git" = "https://github.com/lance-format/lance.git" }
+lance-datagen = { "version" = "=8.0.0-beta.9", "tag" = "v8.0.0-beta.9", "git" = "https://github.com/lance-format/lance.git" }
+lance-file = { "version" = "=8.0.0-beta.9", "tag" = "v8.0.0-beta.9", "git" = "https://github.com/lance-format/lance.git" }
+lance-io = { "version" = "=8.0.0-beta.9", default-features = false, "tag" = "v8.0.0-beta.9", "git" = "https://github.com/lance-format/lance.git" }
+lance-index = { "version" = "=8.0.0-beta.9", "tag" = "v8.0.0-beta.9", "git" = "https://github.com/lance-format/lance.git" }
+lance-linalg = { "version" = "=8.0.0-beta.9", "tag" = "v8.0.0-beta.9", "git" = "https://github.com/lance-format/lance.git" }
+lance-namespace = { "version" = "=8.0.0-beta.9", "tag" = "v8.0.0-beta.9", "git" = "https://github.com/lance-format/lance.git" }
+lance-namespace-impls = { "version" = "=8.0.0-beta.9", default-features = false, "tag" = "v8.0.0-beta.9", "git" = "https://github.com/lance-format/lance.git" }
+lance-table = { "version" = "=8.0.0-beta.9", "tag" = "v8.0.0-beta.9", "git" = "https://github.com/lance-format/lance.git" }
+lance-testing = { "version" = "=8.0.0-beta.9", "tag" = "v8.0.0-beta.9", "git" = "https://github.com/lance-format/lance.git" }
+lance-datafusion = { "version" = "=8.0.0-beta.9", "tag" = "v8.0.0-beta.9", "git" = "https://github.com/lance-format/lance.git" }
+lance-encoding = { "version" = "=8.0.0-beta.9", "tag" = "v8.0.0-beta.9", "git" = "https://github.com/lance-format/lance.git" }
+lance-arrow = { "version" = "=8.0.0-beta.9", "tag" = "v8.0.0-beta.9", "git" = "https://github.com/lance-format/lance.git" }
 ahash = "0.8"
 # Note that this one does not include pyarrow
 arrow = { version = "58.0.0", optional = false }

REVIEW.md

@@ -0,0 +1,26 @@
+# Code review guidelines
+
+Repo-specific guidance for automated PR reviews.
+
+## Cross-SDK parity
+
+LanceDB exposes the same core (`rust/lancedb`) through Python, TypeScript (`nodejs`),
+and Java bindings. Behavioral drift between SDKs is a recurring problem, so watch for
+parity gaps when reviewing — but only flag real ones:
+
+* If the change adds or modifies user-facing API or behavior in the shared core
+  (`rust/lancedb`), check whether each binding that should expose it (`python`,
+  `nodejs`) does. A core change with no corresponding binding update is worth a note.
+* If the change adds or modifies a public API in one SDK but not the other, open the
+  sibling SDK's corresponding module and state whether an equivalent exists. If not,
+  note it as a possible parity gap and suggest a follow-up issue.
+* For bug fixes, first read the sibling SDK's analogous code path to check whether the
+  same bug exists there. Only raise parity if it actually does. Do not ask to "port" a
+  fix for a bug that only ever existed in one binding.
+* Stay silent on internal-only refactors, tests, docs, and changes with no cross-SDK
+  surface.
+* Parity expectations apply to the Python and TypeScript (`nodejs`) SDKs. Java currently
+  implements only the remote table, not the local/embedded backend, so it is expected to
+  be partial — do not flag Java for missing local-only functionality.
+* Keep parity feedback to a short, clearly-labeled note (e.g. "Possible SDK parity
+  gap: …"). It is advisory, not a merge blocker.

ci/check_lance_release.py

@@ -112,25 +112,25 @@ def fetch_remote_tags() -> List[TagInfo]:
             "api",
             "-X",
             "GET",
-            f"repos/{LANCE_REPO}/git/refs/tags",
-            "--paginate",
+            f"repos/{LANCE_REPO}/releases",
             "--jq",
-            ".[].ref",
+            ".[].tag_name",
+            "-F",
+            "per_page=20",
         ]
     )
     tags: List[TagInfo] = []
     for line in output.splitlines():
-        ref = line.strip()
-        if not ref.startswith("refs/tags/v"):
+        tag = line.strip()
+        if not tag.startswith("v"):
             continue
-        tag = ref.split("refs/tags/")[-1]
         version = tag.lstrip("v")
         try:
             tags.append(TagInfo(tag=tag, version=version, semver=parse_semver(version)))
         except ValueError:
             continue
     if not tags:
-        raise RuntimeError("No Lance tags could be parsed from GitHub API output")
+        raise RuntimeError("No Lance releases could be parsed from GitHub API output")
     return tags
 
 

ci/update_lance_dependency.py

@@ -0,0 +1,126 @@
+#!/usr/bin/env python3
+"""Prepare a Lance dependency update for LanceDB."""
+
+from __future__ import annotations
+
+import argparse
+import json
+import re
+import subprocess
+import sys
+from pathlib import Path
+from typing import Sequence
+
+try:
+    from check_lance_release import parse_semver
+except ModuleNotFoundError:
+    # Supports importing as ci.update_lance_dependency from tests or ad hoc checks.
+    from ci.check_lance_release import parse_semver  # type: ignore
+
+
+def normalize_version(raw: str) -> str:
+    value = raw.strip()
+    value = value.removeprefix("refs/tags/")
+    value = value.removeprefix("v")
+    try:
+        parse_semver(value)
+    except ValueError:
+        raise ValueError(f"Unsupported Lance version or tag: {raw}")
+    return value
+
+
+def normalized_tag(version: str) -> str:
+    return f"v{version}"
+
+
+def branch_name(version: str) -> str:
+    suffix = re.sub(r"[^a-zA-Z0-9]+", "-", version).strip("-")
+    suffix = re.sub(r"-+", "-", suffix)
+    return f"codex/update-lance-{suffix}"
+
+
+def commit_type(version: str) -> str:
+    prerelease = version.split("-", maxsplit=1)[1] if "-" in version else ""
+    return "chore" if "beta" in prerelease or "rc" in prerelease else "feat"
+
+
+def metadata_for(version: str) -> dict[str, str]:
+    kind = commit_type(version)
+    message = f"{kind}: update lance dependency to v{version}"
+    return {
+        "version": version,
+        "tag": normalized_tag(version),
+        "branch_name": branch_name(version),
+        "commit_type": kind,
+        "commit_message": message,
+        "pr_title": message,
+    }
+
+
+def run_command(cmd: Sequence[str], *, cwd: Path) -> None:
+    subprocess.run(cmd, cwd=cwd, check=True)
+
+
+def update_java_lance_core_version(repo_root: Path, version: str) -> None:
+    pom_path = repo_root / "java" / "pom.xml"
+    contents = pom_path.read_text(encoding="utf-8")
+    updated, count = re.subn(
+        r"(<lance-core\.version>)[^<]+(</lance-core\.version>)",
+        rf"\g<1>{version}\g<2>",
+        contents,
+        count=1,
+    )
+    if count != 1:
+        raise RuntimeError(
+            "Expected exactly one <lance-core.version> entry in java/pom.xml"
+        )
+    pom_path.write_text(updated, encoding="utf-8")
+
+
+def write_github_outputs(path: str | None, payload: dict[str, str]) -> None:
+    if not path:
+        return
+    with open(path, "a", encoding="utf-8") as output:
+        for key, value in payload.items():
+            output.write(f"{key}={value}\n")
+
+
+def main(argv: Sequence[str] | None = None) -> int:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument(
+        "tag_or_version",
+        help="Lance tag or version, for example refs/tags/v7.2.0-beta.1 or 7.2.0",
+    )
+    parser.add_argument(
+        "--repo-root",
+        type=Path,
+        default=Path(__file__).resolve().parents[1],
+        help="Path to the lancedb repository root",
+    )
+    parser.add_argument(
+        "--github-output",
+        default=None,
+        help="Optional GitHub Actions output file to receive metadata fields",
+    )
+    parser.add_argument(
+        "--metadata-only",
+        action="store_true",
+        help="Only print derived metadata; do not modify dependency files",
+    )
+    args = parser.parse_args(argv)
+
+    repo_root = args.repo_root.resolve()
+    version = normalize_version(args.tag_or_version)
+    payload = metadata_for(version)
+
+    if not args.metadata_only:
+        run_command([sys.executable, "ci/set_lance_version.py", version], cwd=repo_root)
+        update_java_lance_core_version(repo_root, version)
+
+    write_github_outputs(args.github_output, payload)
+    print(json.dumps(payload, sort_keys=True))
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())

deny.toml

@@ -147,6 +147,14 @@ allow = [
     "CDLA-Permissive-2.0",
 ]
 confidence-threshold = 0.8
+# Per-crate license exceptions: allow a license for a specific crate only,
+# rather than globally via the `allow` list above.
+exceptions = [
+    # CDDL-1.0 (copyleft) is pulled in only as a dev/profiling dependency via
+    # `inferno` -> `pprof` -> `lance-testing`; it is a test dependency that we
+    # do not distribute, so scope the allowance to `inferno` alone.
+    { allow = ["CDDL-1.0"], crate = "inferno" },
+]
 # Crates whose license cannot be determined from Cargo metadata but whose
 # license we've manually confirmed from upstream. Keep this list minimal.
 [[licenses.clarify]]

docs/src/java/java.md

@@ -14,7 +14,7 @@ Add the following dependency to your `pom.xml`:
 <dependency>
     <groupId>com.lancedb</groupId>
     <artifactId>lancedb-core</artifactId>
-    <version>0.30.0-beta.0</version>
+    <version>0.30.1-beta.2</version>
 </dependency>
 ```
 

docs/src/js/classes/BranchContents.md

@@ -0,0 +1,43 @@
+[**@lancedb/lancedb**](../README.md) • **Docs**
+
+***
+
+[@lancedb/lancedb](../globals.md) / BranchContents
+
+# Class: BranchContents
+
+## Constructors
+
+### new BranchContents()
+
+```ts
+new BranchContents(): BranchContents
+```
+
+#### Returns
+
+[`BranchContents`](BranchContents.md)
+
+## Properties
+
+### manifestSize
+
+```ts
+manifestSize: number;
+```
+
+***
+
+### parentBranch?
+
+```ts
+optional parentBranch: string;
+```
+
+***
+
+### parentVersion
+
+```ts
+parentVersion: number;
+```

docs/src/js/classes/Branches.md

@@ -0,0 +1,96 @@
+[**@lancedb/lancedb**](../README.md) • **Docs**
+
+***
+
+[@lancedb/lancedb](../globals.md) / Branches
+
+# Class: Branches
+
+Branch manager for a [Table](Table.md).
+
+Unlike tags, `create` and `checkout` return a new [Table](Table.md) handle scoped
+to the branch; writes on it do not affect `main`.
+
+## Methods
+
+### checkout()
+
+```ts
+checkout(name, version?): Promise<Table>
+```
+
+Check out an existing branch and return a handle scoped to it.
+
+With `version` set, the returned handle is pinned to that version of the
+branch (a read-only, detached view); otherwise it tracks the branch's
+latest and stays writable.
+
+#### Parameters
+
+* **name**: `string`
+
+* **version?**: `number`
+
+#### Returns
+
+`Promise`&lt;[`Table`](Table.md)&gt;
+
+***
+
+### create()
+
+```ts
+create(
+   name,
+   fromRef?,
+   fromVersion?): Promise<Table>
+```
+
+Create a branch and return a handle scoped to it.
+
+#### Parameters
+
+* **name**: `string`
+    Name of the new branch.
+
+* **fromRef?**: `string`
+    Source branch to fork from. Defaults to `main`.
+
+* **fromVersion?**: `number`
+    A specific version on `fromRef`. Defaults to latest.
+
+#### Returns
+
+`Promise`&lt;[`Table`](Table.md)&gt;
+
+***
+
+### delete()
+
+```ts
+delete(name): Promise<void>
+```
+
+Delete a branch.
+
+#### Parameters
+
+* **name**: `string`
+
+#### Returns
+
+`Promise`&lt;`void`&gt;
+
+***
+
+### list()
+
+```ts
+list(): Promise<Record<string, BranchContents>>
+```
+
+List all branches, mapping name to branch metadata.
+
+#### Returns
+
+`Promise`&lt;`Record`&lt;`string`, [`BranchContents`](BranchContents.md)&gt;&gt;

docs/src/js/classes/MergeInsertBuilder.md

@@ -76,6 +76,57 @@ the query optimizer chooses a suboptimal path.
 
 ***
 
+### useLsmWrite()
+
+```ts
+useLsmWrite(useLsmWrite): MergeInsertBuilder
+```
+
+Controls whether the merge uses the MemWAL LSM write path.
+
+By default (unset), a `mergeInsert` on a table with an LSM write spec is
+routed through Lance's MemWAL shard writer, and a table without one uses
+the standard path. Pass `false` to force the standard path even when a
+spec is set. Pass `true` to require a spec — `mergeInsert` rejects if none
+is installed.
+
+#### Parameters
+
+* **useLsmWrite**: `boolean`
+    Whether to use the LSM write path.
+
+#### Returns
+
+[`MergeInsertBuilder`](MergeInsertBuilder.md)
+
+***
+
+### validateSingleShard()
+
+```ts
+validateSingleShard(validateSingleShard): MergeInsertBuilder
+```
+
+Controls how an LSM merge checks that its input targets a single shard.
+
+When a table has an LSM write spec, every row in a `mergeInsert` call must
+route to the same shard. When `true` (the default), every row is inspected
+to verify this. When `false`, only the first row is inspected and the
+shard it routes to is used for the whole input — a faster path for callers
+that have already pre-sharded their input. Has no effect on tables without
+an LSM write spec.
+
+#### Parameters
+
+* **validateSingleShard**: `boolean`
+    Whether to check every row routes to one shard. Defaults to `true`.
+
+#### Returns
+
+[`MergeInsertBuilder`](MergeInsertBuilder.md)
+
+***
+
 ### whenMatchedUpdateAll()
 
 ```ts

docs/src/js/classes/Table.md

@@ -110,6 +110,23 @@ containing the new version number of the table after altering the columns.
 
 ***
 
+### branches()
+
+```ts
+abstract branches(): Promise<Branches>
+```
+
+Get the branch manager for this table.
+
+Branches are isolated, writable lines of history forked from another
+branch (or version). Writes on a branch do not affect `main`.
+
+#### Returns
+
+`Promise`&lt;[`Branches`](Branches.md)&gt;
+
+***
+
 ### checkout()
 
 ```ts
@@ -187,6 +204,25 @@ Any attempt to use the table after it is closed will result in an error.
 
 ***
 
+### closeLsmWriters()
+
+```ts
+abstract closeLsmWriters(): Promise<void>
+```
+
+Drain and close any cached MemWAL shard writers held for this table.
+
+When an [LsmWriteSpec](../interfaces/LsmWriteSpec.md) is installed, `mergeInsert` opens MemWAL
+shard writers and caches them for reuse across calls. This closes them,
+flushing pending data; writers reopen lazily on the next `mergeInsert`.
+It is a no-op when no writers are cached.
+
+#### Returns
+
+`Promise`&lt;`void`&gt;
+
+***
+
 ### countRows()
 
 ```ts
@@ -975,6 +1011,29 @@ based on the row being updated (e.g. "my_col + 1")
 
 ***
 
+### updateFieldMetadata()
+
+```ts
+abstract updateFieldMetadata(updates): Promise<UpdateFieldMetadataResult>
+```
+
+Update per-field (column) metadata.
+
+#### Parameters
+
+* **updates**: [`FieldMetadataUpdate`](../interfaces/FieldMetadataUpdate.md)[]
+    One or more per-field updates. Each
+    update's metadata is merged into the field's existing metadata by default;
+    a value of `null` deletes that key, and `replace: true` swaps the whole map.
+
+#### Returns
+
+`Promise`&lt;[`UpdateFieldMetadataResult`](../interfaces/UpdateFieldMetadataResult.md)&gt;
+
+resolves to the new table version.
+
+***
+
 ### vectorSearch()
 
 ```ts

docs/src/js/globals.md

@@ -19,6 +19,8 @@
 
 - [BooleanQuery](classes/BooleanQuery.md)
 - [BoostQuery](classes/BoostQuery.md)
+- [BranchContents](classes/BranchContents.md)
+- [Branches](classes/Branches.md)
 - [Connection](classes/Connection.md)
 - [HeaderProvider](classes/HeaderProvider.md)
 - [Index](classes/Index.md)
@@ -65,6 +67,7 @@
 - [DropNamespaceOptions](interfaces/DropNamespaceOptions.md)
 - [DropNamespaceResponse](interfaces/DropNamespaceResponse.md)
 - [ExecutableQuery](interfaces/ExecutableQuery.md)
+- [FieldMetadataUpdate](interfaces/FieldMetadataUpdate.md)
 - [FragmentStatistics](interfaces/FragmentStatistics.md)
 - [FragmentSummaryStats](interfaces/FragmentSummaryStats.md)
 - [FtsOptions](interfaces/FtsOptions.md)
@@ -101,6 +104,7 @@
 - [TimeoutConfig](interfaces/TimeoutConfig.md)
 - [TlsConfig](interfaces/TlsConfig.md)
 - [TokenResponse](interfaces/TokenResponse.md)
+- [UpdateFieldMetadataResult](interfaces/UpdateFieldMetadataResult.md)
 - [UpdateOptions](interfaces/UpdateOptions.md)
 - [UpdateResult](interfaces/UpdateResult.md)
 - [Version](interfaces/Version.md)

docs/src/js/interfaces/ConnectionOptions.md

@@ -70,16 +70,20 @@ client used by manifest-enabled native connections.
 optional readConsistencyInterval: number;
 ```
 
-(For LanceDB OSS only): The interval, in seconds, at which to check for
-updates to the table from other processes. If None, then consistency is not
-checked. For performance reasons, this is the default. For strong
-consistency, set this to zero seconds. Then every read will check for
-updates from other processes. As a compromise, you can set this to a
-non-zero value for eventual consistency. If more than that interval
-has passed since the last check, then the table will be checked for updates.
-Note: this consistency only applies to read operations. Write operations are
+The interval, in seconds, at which to check for updates to the table
+from other processes. If None, then consistency is not checked. For
+performance reasons, this is the default. For strong consistency, set
+this to zero seconds. Then every read will check for updates from other
+processes. As a compromise, you can set this to a non-zero value for
+eventual consistency. If more than that interval has passed since the
+last check, then the table will be checked for updates. Note: this
+consistency only applies to read operations. Write operations are
 always consistent.
 
+Stronger consistency is not free. The smaller the interval, the more
+often each read pays the cost of checking for updates against object
+storage, raising per-read latency and cost.
+
 ***
 
 ### region?

docs/src/js/interfaces/FieldMetadataUpdate.md

@@ -0,0 +1,41 @@
+[**@lancedb/lancedb**](../README.md) • **Docs**
+
+***
+
+[@lancedb/lancedb](../globals.md) / FieldMetadataUpdate
+
+# Interface: FieldMetadataUpdate
+
+A per-field metadata update, addressed by dot-path.
+
+## Properties
+
+### metadata
+
+```ts
+metadata: Record<string, null | string>;
+```
+
+Metadata key/value pairs. Merged into the field's existing metadata by
+default; a value of `null` deletes that key.
+
+***
+
+### path
+
+```ts
+path: string;
+```
+
+Dot-separated path to the field. For a top-level column this is just its
+name; for a nested field it's the path, e.g. "a.b.c".
+
+***
+
+### replace?
+
+```ts
+optional replace: boolean;
+```
+
+If true, replace the field's entire metadata map instead of merging.

docs/src/js/interfaces/IndexStatistics.md

@@ -30,17 +30,6 @@ The type of the index
 
 ***
 
-### loss?
-
-```ts
-optional loss: number;
-```
-
-The KMeans loss value of the index,
-it is only present for vector indices.
-
-***
-
 ### numIndexedRows
 
 ```ts

docs/src/js/interfaces/LsmWriteSpec.md

@@ -11,7 +11,10 @@ Specification selecting Lance's MemWAL LSM-style write path for
 
 `specType` is `"bucket"`, `"identity"`, or `"unsharded"`. For `"bucket"`,
 `column` and `numBuckets` are required; for `"identity"`, `column` is
-required.
+required and must be a deterministic function of the unenforced primary
+key (every row with a given primary key must always produce the same
+`column` value, or upserts of that key can land in different shards and a
+stale version can win).
 
 ## Properties
 

docs/src/js/interfaces/MergeResult.md

@@ -32,6 +32,14 @@ numInsertedRows: number;
 
 ***
 
+### numRows
+
+```ts
+numRows: number;
+```
+
+***
+
 ### numUpdatedRows
 
 ```ts

docs/src/js/interfaces/OpenTableOptions.md

@@ -8,6 +8,18 @@
 
 ## Properties
 
+### branch?
+
+```ts
+optional branch: string;
+```
+
+Open the table scoped to this branch instead of the default branch.
+
+Reads and writes on the returned table operate in the branch's context.
+
+***
+
 ### ~~indexCacheSize?~~
 
 ```ts
@@ -43,3 +55,17 @@ Options already set on the connection will be inherited by the table,
 but can be overridden here.
 
 The available options are described at https://docs.lancedb.com/storage/
+
+***
+
+### version?
+
+```ts
+optional version: number;
+```
+
+Open the table pinned to this version, producing a read-only view.
+
+Composes with [OpenTableOptions.branch](OpenTableOptions.md#branch): when both are set, opens
+that branch at the version; otherwise opens `main` at the version. Call
+`checkoutLatest` to return to a writable state.

docs/src/js/interfaces/UpdateFieldMetadataResult.md

@@ -0,0 +1,15 @@
+[**@lancedb/lancedb**](../README.md) • **Docs**
+
+***
+
+[@lancedb/lancedb](../globals.md) / UpdateFieldMetadataResult
+
+# Interface: UpdateFieldMetadataResult
+
+## Properties
+
+### version
+
+```ts
+version: number;
+```

java/lancedb-core/pom.xml

@@ -8,7 +8,7 @@
     <parent>
       <groupId>com.lancedb</groupId>
       <artifactId>lancedb-parent</artifactId>
-      <version>0.30.0-beta.0</version>
+      <version>0.30.1-beta.2</version>
       <relativePath>../pom.xml</relativePath>
     </parent>
 

java/pom.xml

@@ -6,7 +6,7 @@
 
     <groupId>com.lancedb</groupId>
     <artifactId>lancedb-parent</artifactId>
-    <version>0.30.0-beta.0</version>
+    <version>0.30.1-beta.2</version>
     <packaging>pom</packaging>
     <name>${project.artifactId}</name>
     <description>LanceDB Java SDK Parent POM</description>
@@ -28,7 +28,7 @@
     <properties>
         <project.build.sourceEncoding>UTF-8</project.build.sourceEncoding>
         <arrow.version>15.0.0</arrow.version>
-        <lance-core.version>7.0.0-beta.13</lance-core.version>
+        <lance-core.version>8.0.0-beta.9</lance-core.version>
         <spotless.skip>false</spotless.skip>
         <spotless.version>2.30.0</spotless.version>
         <spotless.java.googlejavaformat.version>1.7</spotless.java.googlejavaformat.version>

nodejs/Cargo.toml

@@ -1,7 +1,7 @@
 [package]
 name = "lancedb-nodejs"
 edition.workspace = true
-version = "0.30.0-beta.0"
+version = "0.30.1-beta.2"
 publish = false
 license.workspace = true
 description.workspace = true

nodejs/__test__/connection.test.ts

@@ -171,18 +171,22 @@ describe("given a connection", () => {
 
     let manifestDir =
       tmpDir.name + "/test_manifest_paths_v2_empty.lance/_versions";
-    readdirSync(manifestDir).forEach((file) => {
-      expect(file).toMatch(/^\d{20}\.manifest$/);
-    });
+    readdirSync(manifestDir)
+      .filter((f) => f.endsWith(".manifest"))
+      .forEach((file) => {
+        expect(file).toMatch(/^\d{20}\.manifest$/);
+      });
 
     table = (await db.createTable("test_manifest_paths_v2", [{ id: 1 }], {
       enableV2ManifestPaths: true,
     })) as LocalTable;
     expect(await table.usesV2ManifestPaths()).toBe(true);
     manifestDir = tmpDir.name + "/test_manifest_paths_v2.lance/_versions";
-    readdirSync(manifestDir).forEach((file) => {
-      expect(file).toMatch(/^\d{20}\.manifest$/);
-    });
+    readdirSync(manifestDir)
+      .filter((f) => f.endsWith(".manifest"))
+      .forEach((file) => {
+        expect(file).toMatch(/^\d{20}\.manifest$/);
+      });
   });
 
   it("should be able to migrate tables to the V2 manifest paths", async () => {
@@ -199,16 +203,20 @@ describe("given a connection", () => {
 
     const manifestDir =
       tmpDir.name + "/test_manifest_path_migration.lance/_versions";
-    readdirSync(manifestDir).forEach((file) => {
-      expect(file).toMatch(/^\d\.manifest$/);
-    });
+    readdirSync(manifestDir)
+      .filter((f) => f.endsWith(".manifest"))
+      .forEach((file) => {
+        expect(file).toMatch(/^\d\.manifest$/);
+      });
 
     await table.migrateManifestPathsV2();
     expect(await table.usesV2ManifestPaths()).toBe(true);
 
-    readdirSync(manifestDir).forEach((file) => {
-      expect(file).toMatch(/^\d{20}\.manifest$/);
-    });
+    readdirSync(manifestDir)
+      .filter((f) => f.endsWith(".manifest"))
+      .forEach((file) => {
+        expect(file).toMatch(/^\d{20}\.manifest$/);
+      });
   });
 });
 

nodejs/__test__/remote.test.ts

@@ -191,6 +191,34 @@ describe("remote connection", () => {
     );
   });
 
+  it("allows version on remote but rejects a non-main branch", async () => {
+    await withMockDatabase(
+      (_req, res) => {
+        // describe (table open + version validation) always succeeds
+        const body = JSON.stringify({
+          name: "t",
+          version: 2,
+          schema: { fields: [] },
+        });
+        res.writeHead(200, { "Content-Type": "application/json" }).end(body);
+      },
+      async (db) => {
+        // version-only (and "main" + version) is allowed: remote supports
+        // version time-travel even though it has no branches
+        await db.openTable("t", undefined, { version: 2 });
+        await db.openTable("t", undefined, { branch: "main", version: 2 });
+
+        // a non-main branch is rejected, with or without a version
+        await expect(
+          db.openTable("t", undefined, { branch: "exp" }),
+        ).rejects.toThrow(/branching/);
+        await expect(
+          db.openTable("t", undefined, { branch: "exp", version: 2 }),
+        ).rejects.toThrow(/branching/);
+      },
+    );
+  });
+
   describe("TlsConfig", () => {
     it("should create TlsConfig with all fields", () => {
       const tlsConfig: TlsConfig = {

nodejs/__test__/table.test.ts

@@ -85,6 +85,136 @@ describe.each([arrow15, arrow16, arrow17, arrow18])(
       await expect(table.countRows()).resolves.toBe(3);
     });
 
+    it("should support branches", async () => {
+      await table.add([{ id: 1 }]);
+      expect(await table.countRows()).toBe(1);
+
+      // fork an isolated, writable branch from main
+      const branch = await (await table.branches()).create("exp");
+      expect(await branch.countRows()).toBe(1);
+      await branch.add([{ id: 2 }]);
+      expect(await branch.countRows()).toBe(2);
+      // main is untouched by branch writes
+      expect(await table.countRows()).toBe(1);
+
+      // listed, with main (null) as the parent
+      const list = await (await table.branches()).list();
+      expect(Object.keys(list)).toContain("exp");
+      expect(list["exp"].parentBranch).toBeNull();
+
+      // fromRef="main" is equivalent to the default
+      await (await table.branches()).create("exp2", "main");
+      const list2 = await (await table.branches()).list();
+      expect(list2["exp2"].parentBranch).toBeNull();
+
+      // checkout returns a handle scoped to the branch's latest
+      const checkedOut = await (await table.branches()).checkout("exp");
+      expect(await checkedOut.countRows()).toBe(2);
+
+      // delete removes it
+      await (await table.branches()).delete("exp");
+      await (await table.branches()).delete("exp2");
+      const after = await (await table.branches()).list();
+      expect(Object.keys(after)).not.toContain("exp");
+    });
+
+    it("should open a branch via open_table", async () => {
+      const db = await connect(tmpDir.name);
+      await table.add([{ id: 1 }]);
+      const branch = await (await table.branches()).create("exp");
+      await branch.add([{ id: 2 }]);
+
+      // open_table(..., { branch }) returns a handle scoped to the branch
+      const opened = await db.openTable("some_table", undefined, {
+        branch: "exp",
+      });
+      expect(await opened.countRows()).toBe(2);
+      // opening without branch still tracks main
+      expect(await (await db.openTable("some_table")).countRows()).toBe(1);
+    });
+
+    it("should open a branch at a version isolated from main and HEAD", async () => {
+      const db = await connect(tmpDir.name);
+      // main: a single fork-point row
+      const t = await db.createTable("bv_table", [{ id: 0 }]);
+      const mainV1 = await t.version();
+
+      // fork "exp", then advance exp AND main independently past the fork so
+      // they diverge while sharing version numbers
+      const exp = await (await t.branches()).create("exp");
+      await exp.add([{ id: 1 }]); // exp: {0, 1}
+      const expV2 = await exp.version();
+      await exp.add([{ id: 2 }]); // exp HEAD: {0, 1, 2}
+      await t.add([{ id: 100 }, { id: 101 }, { id: 102 }]); // main HEAD: {0,100,101,102}
+      expect(await t.version()).toBe(expV2);
+
+      // open exp at the shared version: the data must be exp's, not main's.
+      // count alone cannot prove this (main@v2 also exists), so assert
+      // provenance by content.
+      const pinned = await db.openTable("bv_table", undefined, {
+        branch: "exp",
+        version: expV2,
+      });
+      expect(await pinned.countRows()).toBe(2); // not exp HEAD (3), not main@v2 (4)
+      expect(await pinned.countRows("id = 1")).toBe(1); // exp's post-fork row
+      expect(await pinned.countRows("id = 100")).toBe(0); // main's rows invisible
+
+      // the same coordinate is reachable directly via branches().checkout(name, version)
+      const pinnedDirect = await (await t.branches()).checkout("exp", expV2);
+      expect(await pinnedDirect.countRows()).toBe(2);
+
+      // the HEADs are unaffected
+      expect(
+        await (
+          await db.openTable("bv_table", undefined, { branch: "exp" })
+        ).countRows(),
+      ).toBe(3);
+      expect(await (await db.openTable("bv_table")).countRows()).toBe(4);
+
+      // version-only (no branch) time-travels main itself: its fork-point
+      // version holds only main's first row, and the shared version number
+      // resolves to main's data, not the branch's ("opens main at the version")
+      const oldMain = await db.openTable("bv_table", undefined, {
+        version: mainV1,
+      });
+      expect(await oldMain.countRows()).toBe(1);
+      const sharedOnMain = await db.openTable("bv_table", undefined, {
+        version: expV2,
+      });
+      expect(await sharedOnMain.countRows()).toBe(4); // main@v2, not exp@v2 (2)
+
+      // detached head: writing to a pinned version is rejected
+      await expect(pinned.add([{ id: 9 }])).rejects.toThrow(
+        /cannot be modified/,
+      );
+
+      // a nonexistent version is rejected -- on main, and on a branch (a
+      // distinct resolution path, on the branch's manifests)
+      await expect(
+        db.openTable("bv_table", undefined, { version: 9999 }),
+      ).rejects.toThrow();
+      await expect(
+        db.openTable("bv_table", undefined, { branch: "exp", version: 9999 }),
+      ).rejects.toThrow();
+
+      // checkoutLatest re-attaches the pinned handle to the BRANCH's HEAD
+      // (writable again), not main's HEAD (4), and not staying pinned (2)
+      await pinned.checkoutLatest();
+      expect(await pinned.countRows()).toBe(3); // exp HEAD
+      await pinned.add([{ id: 3 }]);
+      expect(await pinned.countRows()).toBe(4); // writable again
+    });
+
+    it("rejects invalid branch inputs", async () => {
+      const branches = await table.branches();
+      await expect(branches.create("")).rejects.toThrow("non-empty");
+      await expect(branches.checkout("")).rejects.toThrow("non-empty");
+      await expect(branches.delete("")).rejects.toThrow("non-empty");
+      await expect(branches.create("bad", "main", -1)).rejects.toThrow(
+        "non-negative",
+      );
+    });
+
     it("should show table stats", async () => {
       await table.add([{ id: 1 }, { id: 2 }]);
       await table.add([{ id: 1 }]);
@@ -721,7 +851,7 @@ describe("When creating an index", () => {
       columns: ["vec"],
     });
     const stats = await tbl.indexStats("vec_idx");
-    expect(stats?.loss).toBeDefined();
+    expect(stats).toBeDefined();
 
     // Search without specifying the column
     let rst = await tbl
@@ -1150,7 +1280,6 @@ describe("When creating an index", () => {
     expect(stats?.distanceType).toBeUndefined();
     expect(stats?.indexType).toEqual("BTREE");
     expect(stats?.numIndices).toEqual(1);
-    expect(stats?.loss).toBeUndefined();
   });
 
   test("when getting stats on non-existent index", async () => {
@@ -1571,6 +1700,33 @@ describe("schema evolution", function () {
     expect(await table.schema()).toEqual(expectedSchema3);
   });
 
+  it("can update field metadata", async function () {
+    const con = await connect(tmpDir.name);
+    const table = await con.createTable("fm", [
+      { id: 1, category: "a" },
+      { id: 2, category: "b" },
+    ]);
+
+    const res = await table.updateFieldMetadata([
+      { path: "category", metadata: { unit: "label", pii: "false" } },
+    ]);
+    expect(res).toHaveProperty("version");
+    expect(res.version).toBe(2);
+
+    let cat = (await table.schema()).fields.find((f) => f.name === "category");
+    expect(cat?.metadata.get("unit")).toBe("label");
+    expect(cat?.metadata.get("pii")).toBe("false");
+
+    // merge: add a key, delete one via null, keep the rest
+    await table.updateFieldMetadata([
+      { path: "category", metadata: { source: "import", pii: null } },
+    ]);
+    cat = (await table.schema()).fields.find((f) => f.name === "category");
+    expect(cat?.metadata.get("unit")).toBe("label"); // preserved
+    expect(cat?.metadata.get("source")).toBe("import"); // added
+    expect(cat?.metadata.has("pii")).toBe(false); // deleted
+  });
+
   it("can cast to various types", async function () {
     const con = await connect(tmpDir.name);
 
@@ -2625,3 +2781,97 @@ describe("setLsmWriteSpec / unsetLsmWriteSpec", () => {
     ).rejects.toThrow();
   });
 });
+
+describe("LSM merge insert", () => {
+  let tmpDir: tmp.DirResult;
+
+  beforeEach(() => {
+    tmpDir = tmp.dirSync({ unsafeCleanup: true });
+  });
+  afterEach(() => tmpDir.removeCallback());
+
+  async function bucketTable(conn: Connection): Promise<Table> {
+    // The primary key column must be non-nullable.
+    const table = await conn.createEmptyTable(
+      "t",
+      new arrow.Schema([
+        new arrow.Field("id", new arrow.Utf8(), false),
+        new arrow.Field("value", new arrow.Float64(), true),
+      ]),
+    );
+    await table.add([
+      { id: "a", value: 1 },
+      { id: "b", value: 2 },
+    ]);
+    await table.setUnenforcedPrimaryKey("id");
+    // numBuckets = 1: every row routes to the single bucket.
+    await table.setLsmWriteSpec({
+      specType: "bucket",
+      column: "id",
+      numBuckets: 1,
+    });
+    return table;
+  }
+
+  it("routes merge_insert through the shard writer", async () => {
+    const conn = await connect(tmpDir.name);
+    const table = await bucketTable(conn);
+
+    const res = await table
+      .mergeInsert("id")
+      .whenMatchedUpdateAll()
+      .whenNotMatchedInsertAll()
+      .execute([
+        { id: "c", value: 3 },
+        { id: "d", value: 4 },
+      ]);
+    // LSM path: rows go to the MemWAL, so only numRows is populated.
+    expect(res.numRows).toBe(2);
+    expect(res.version).toBe(0);
+    expect(res.numInsertedRows).toBe(0);
+
+    await table.closeLsmWriters();
+  });
+
+  it("falls back to the standard path with useLsmWrite(false)", async () => {
+    const conn = await connect(tmpDir.name);
+    const table = await bucketTable(conn);
+
+    const res = await table
+      .mergeInsert("id")
+      .whenNotMatchedInsertAll()
+      .useLsmWrite(false)
+      .execute([
+        { id: "b", value: 9 },
+        { id: "e", value: 5 },
+      ]);
+    // Standard path commits: id="e" inserted ("b" already exists).
+    expect(res.numInsertedRows).toBe(1);
+    expect(await table.countRows()).toBe(3);
+  });
+
+  it("supports validateSingleShard(false)", async () => {
+    const conn = await connect(tmpDir.name);
+    const table = await bucketTable(conn);
+
+    const res = await table
+      .mergeInsert("id")
+      .whenMatchedUpdateAll()
+      .whenNotMatchedInsertAll()
+      .validateSingleShard(false)
+      .execute([{ id: "f", value: 6 }]);
+    expect(res.numRows).toBe(1);
+  });
+
+  it("rejects a non-upsert merge under an LSM spec", async () => {
+    const conn = await connect(tmpDir.name);
+    const table = await bucketTable(conn);
+
+    await expect(
+      table
+        .mergeInsert("id")
+        .whenNotMatchedInsertAll()
+        .execute([{ id: "g", value: 7 }]),
+    ).rejects.toThrow();
+  });
+});

nodejs/lancedb/connection.ts

@@ -84,6 +84,20 @@ export interface CreateTableOptions {
 }
 
 export interface OpenTableOptions {
+  /**
+   * Open the table scoped to this branch instead of the default branch.
+   *
+   * Reads and writes on the returned table operate in the branch's context.
+   */
+  branch?: string;
+  /**
+   * Open the table pinned to this version, producing a read-only view.
+   *
+   * Composes with {@link OpenTableOptions.branch}: when both are set, opens
+   * that branch at the version; otherwise opens `main` at the version. Call
+   * `checkoutLatest` to return to a writable state.
+   */
+  version?: number;
   /**
    * Configuration for object storage.
    *
@@ -483,7 +497,20 @@ export class LocalConnection extends Connection {
       options?.indexCacheSize,
     );
 
-    return new LocalTable(innerTable);
+    let table: Table = new LocalTable(innerTable);
+    // "main" is the default branch, so treat it as no branch. On a real branch,
+    // scope and pin in one step (yielding "version V of branch B"); otherwise
+    // pin the version, if any, against main.
+    const branch =
+      options?.branch != null && options.branch !== "main"
+        ? options.branch
+        : undefined;
+    if (branch != null) {
+      table = await (await table.branches()).checkout(branch, options?.version);
+    } else if (options?.version != null) {
+      await table.checkout(options.version);
+    }
+    return table;
   }
 
   async cloneTable(

nodejs/lancedb/index.ts

@@ -38,10 +38,12 @@ export {
   FragmentSummaryStats,
   Tags,
   TagContents,
+  BranchContents,
   MergeResult,
   AddResult,
   AddColumnsResult,
   AlterColumnsResult,
+  UpdateFieldMetadataResult,
   DeleteResult,
   DropColumnsResult,
   UpdateResult,
@@ -110,6 +112,7 @@ export {
 
 export {
   Table,
+  Branches,
   AddDataOptions,
   UpdateOptions,
   OptimizeOptions,
@@ -117,6 +120,7 @@ export {
   WriteProgress,
   LsmWriteSpec,
   ColumnAlteration,
+  FieldMetadataUpdate,
 } from "./table";
 
 export {

nodejs/lancedb/merge.ts

@@ -87,6 +87,41 @@ export class MergeInsertBuilder {
       this.#schema,
     );
   }
+  /**
+   * Controls whether the merge uses the MemWAL LSM write path.
+   *
+   * By default (unset), a `mergeInsert` on a table with an LSM write spec is
+   * routed through Lance's MemWAL shard writer, and a table without one uses
+   * the standard path. Pass `false` to force the standard path even when a
+   * spec is set. Pass `true` to require a spec — `mergeInsert` rejects if none
+   * is installed.
+   *
+   * @param useLsmWrite - Whether to use the LSM write path.
+   */
+  useLsmWrite(useLsmWrite: boolean): MergeInsertBuilder {
+    return new MergeInsertBuilder(
+      this.#native.useLsmWrite(useLsmWrite),
+      this.#schema,
+    );
+  }
+  /**
+   * Controls how an LSM merge checks that its input targets a single shard.
+   *
+   * When a table has an LSM write spec, every row in a `mergeInsert` call must
+   * route to the same shard. When `true` (the default), every row is inspected
+   * to verify this. When `false`, only the first row is inspected and the
+   * shard it routes to is used for the whole input — a faster path for callers
+   * that have already pre-sharded their input. Has no effect on tables without
+   * an LSM write spec.
+   *
+   * @param validateSingleShard - Whether to check every row routes to one shard. Defaults to `true`.
+   */
+  validateSingleShard(validateSingleShard: boolean): MergeInsertBuilder {
+    return new MergeInsertBuilder(
+      this.#native.validateSingleShard(validateSingleShard),
+      this.#schema,
+    );
+  }
   /**
    * Executes the merge insert operation
    *

nodejs/lancedb/table.ts

@@ -25,13 +25,16 @@ import {
   AddColumnsSql,
   AddResult,
   AlterColumnsResult,
+  BranchContents,
   DeleteResult,
   DropColumnsResult,
   IndexConfig,
   IndexStatistics,
+  Branches as NativeBranches,
   OptimizeStats,
   TableStatistics,
   Tags,
+  UpdateFieldMetadataResult,
   UpdateResult,
   Table as _NativeTable,
 } from "./native";
@@ -161,7 +164,10 @@ export interface Version {
  *
  * `specType` is `"bucket"`, `"identity"`, or `"unsharded"`. For `"bucket"`,
  * `column` and `numBuckets` are required; for `"identity"`, `column` is
- * required.
+ * required and must be a deterministic function of the unenforced primary
+ * key (every row with a given primary key must always produce the same
+ * `column` value, or upserts of that key can land in different shards and a
+ * stale version can win).
  */
 export interface LsmWriteSpec {
   /** One of `"bucket"`, `"identity"`, or `"unsharded"`. */
@@ -505,6 +511,18 @@ export abstract class Table {
   abstract alterColumns(
     columnAlterations: ColumnAlteration[],
   ): Promise<AlterColumnsResult>;
+
+  /**
+   * Update per-field (column) metadata.
+   * @param {FieldMetadataUpdate[]} updates One or more per-field updates. Each
+   * update's metadata is merged into the field's existing metadata by default;
+   * a value of `null` deletes that key, and `replace: true` swaps the whole map.
+   * @returns {Promise<UpdateFieldMetadataResult>} resolves to the new table version.
+   */
+  abstract updateFieldMetadata(
+    updates: FieldMetadataUpdate[],
+  ): Promise<UpdateFieldMetadataResult>;
+
   /**
    * Drop one or more columns from the dataset
    *
@@ -567,6 +585,16 @@ export abstract class Table {
    * @returns {Promise<void>}
    */
   abstract unsetLsmWriteSpec(): Promise<void>;
+  /**
+   * Drain and close any cached MemWAL shard writers held for this table.
+   *
+   * When an {@link LsmWriteSpec} is installed, `mergeInsert` opens MemWAL
+   * shard writers and caches them for reuse across calls. This closes them,
+   * flushing pending data; writers reopen lazily on the next `mergeInsert`.
+   * It is a no-op when no writers are cached.
+   * @returns {Promise<void>}
+   */
+  abstract closeLsmWriters(): Promise<void>;
   /** Retrieve the version of the table */
 
   abstract version(): Promise<number>;
@@ -627,6 +655,14 @@ export abstract class Table {
    */
   abstract tags(): Promise<Tags>;
 
+  /**
+   * Get the branch manager for this table.
+   *
+   * Branches are isolated, writable lines of history forked from another
+   * branch (or version). Writes on a branch do not affect `main`.
+   */
+  abstract branches(): Promise<Branches>;
+
   /**
    * Restore the table to the currently checked out version
    *
@@ -1024,6 +1060,12 @@ export class LocalTable extends Table {
     return await this.inner.alterColumns(processedAlterations);
   }
 
+  async updateFieldMetadata(
+    updates: FieldMetadataUpdate[],
+  ): Promise<UpdateFieldMetadataResult> {
+    return await this.inner.updateFieldMetadata(updates);
+  }
+
   async dropColumns(columnNames: string[]): Promise<DropColumnsResult> {
     return await this.inner.dropColumns(columnNames);
   }
@@ -1041,6 +1083,10 @@ export class LocalTable extends Table {
     return await this.inner.unsetLsmWriteSpec();
   }
 
+  async closeLsmWriters(): Promise<void> {
+    return await this.inner.closeLsmWriters();
+  }
+
   async version(): Promise<number> {
     return await this.inner.version();
   }
@@ -1072,6 +1118,10 @@ export class LocalTable extends Table {
     return await this.inner.tags();
   }
 
+  async branches(): Promise<Branches> {
+    return new Branches(await this.inner.branches());
+  }
+
   async optimize(options?: Partial<OptimizeOptions>): Promise<OptimizeStats> {
     let cleanupOlderThanMs;
     if (
@@ -1186,3 +1236,73 @@ export interface ColumnAlteration {
   /** Set the new nullability. Note that a nullable column cannot be made non-nullable. */
   nullable?: boolean;
 }
+
+/** A per-field metadata update, addressed by dot-path. */
+export interface FieldMetadataUpdate {
+  /**
+   * Dot-separated path to the field. For a top-level column this is just its
+   * name; for a nested field it's the path, e.g. "a.b.c".
+   */
+  path: string;
+  /**
+   * Metadata key/value pairs. Merged into the field's existing metadata by
+   * default; a value of `null` deletes that key.
+   */
+  metadata: Record<string, string | null>;
+  /** If true, replace the field's entire metadata map instead of merging. */
+  replace?: boolean;
+}
+
+/**
+ * Branch manager for a {@link Table}.
+ *
+ * Unlike tags, `create` and `checkout` return a new {@link Table} handle scoped
+ * to the branch; writes on it do not affect `main`.
+ */
+export class Branches {
+  #inner: NativeBranches;
+
+  /**
+   * Construct a Branches manager. Internal use only.
+   * @hidden
+   */
+  constructor(inner: NativeBranches) {
+    this.#inner = inner;
+  }
+
+  /** List all branches, mapping name to branch metadata. */
+  async list(): Promise<Record<string, BranchContents>> {
+    return await this.#inner.list();
+  }
+
+  /**
+   * Create a branch and return a handle scoped to it.
+   *
+   * @param name Name of the new branch.
+   * @param fromRef Source branch to fork from. Defaults to `main`.
+   * @param fromVersion A specific version on `fromRef`. Defaults to latest.
+   */
+  async create(
+    name: string,
+    fromRef?: string,
+    fromVersion?: number,
+  ): Promise<Table> {
+    return new LocalTable(await this.#inner.create(name, fromRef, fromVersion));
+  }
+
+  /**
+   * Check out an existing branch and return a handle scoped to it.
+   *
+   * With `version` set, the returned handle is pinned to that version of the
+   * branch (a read-only, detached view); otherwise it tracks the branch's
+   * latest and stays writable.
+   */
+  async checkout(name: string, version?: number): Promise<Table> {
+    return new LocalTable(await this.#inner.checkout(name, version));
+  }
+
+  /** Delete a branch. */
+  async delete(name: string): Promise<void> {
+    return await this.#inner.delete(name);
+  }
+}

nodejs/npm/darwin-arm64/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@lancedb/lancedb-darwin-arm64",
-	"version": "0.30.0-beta.0",
+	"version": "0.30.1-beta.2",
 	"os": ["darwin"],
 	"cpu": ["arm64"],
 	"main": "lancedb.darwin-arm64.node",

nodejs/npm/linux-arm64-gnu/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@lancedb/lancedb-linux-arm64-gnu",
-	"version": "0.30.0-beta.0",
+	"version": "0.30.1-beta.2",
 	"os": ["linux"],
 	"cpu": ["arm64"],
 	"main": "lancedb.linux-arm64-gnu.node",

nodejs/npm/linux-arm64-musl/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@lancedb/lancedb-linux-arm64-musl",
-	"version": "0.30.0-beta.0",
+	"version": "0.30.1-beta.2",
 	"os": ["linux"],
 	"cpu": ["arm64"],
 	"main": "lancedb.linux-arm64-musl.node",

nodejs/npm/linux-x64-gnu/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@lancedb/lancedb-linux-x64-gnu",
-	"version": "0.30.0-beta.0",
+	"version": "0.30.1-beta.2",
 	"os": ["linux"],
 	"cpu": ["x64"],
 	"main": "lancedb.linux-x64-gnu.node",

nodejs/npm/linux-x64-musl/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@lancedb/lancedb-linux-x64-musl",
-	"version": "0.30.0-beta.0",
+	"version": "0.30.1-beta.2",
 	"os": ["linux"],
 	"cpu": ["x64"],
 	"main": "lancedb.linux-x64-musl.node",

nodejs/npm/win32-arm64-msvc/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@lancedb/lancedb-win32-arm64-msvc",
-  "version": "0.30.0-beta.0",
+  "version": "0.30.1-beta.2",
   "os": [
     "win32"
   ],

nodejs/npm/win32-x64-msvc/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@lancedb/lancedb-win32-x64-msvc",
-	"version": "0.30.0-beta.0",
+	"version": "0.30.1-beta.2",
 	"os": ["win32"],
 	"cpu": ["x64"],
 	"main": "lancedb.win32-x64-msvc.node",

nodejs/package-lock.json

@@ -1,12 +1,12 @@
 {
   "name": "@lancedb/lancedb",
-  "version": "0.30.0-beta.0",
+  "version": "0.30.1-beta.2",
   "lockfileVersion": 3,
   "requires": true,
   "packages": {
     "": {
       "name": "@lancedb/lancedb",
-      "version": "0.30.0-beta.0",
+      "version": "0.30.1-beta.2",
       "cpu": [
         "x64",
         "arm64"
@@ -26,7 +26,7 @@
         "@aws-sdk/client-s3": "3.1003.0",
         "@biomejs/biome": "^1.7.3",
         "@jest/globals": "^29.7.0",
-        "@napi-rs/cli": "3.5.1",
+        "@napi-rs/cli": "3.7.0",
         "@types/axios": "^0.14.0",
         "@types/jest": "^29.1.2",
         "@types/node": "22.7.4",
@@ -2942,9 +2942,9 @@
       }
     },
     "node_modules/@napi-rs/cli": {
-      "version": "3.5.1",
-      "resolved": "https://registry.npmjs.org/@napi-rs/cli/-/cli-3.5.1.tgz",
-      "integrity": "sha512-XBfLQRDcB3qhu6bazdMJsecWW55kR85l5/k0af9BIBELXQSsCFU0fzug7PX8eQp6vVdm7W/U3z6uP5WmITB2Gw==",
+      "version": "3.7.0",
+      "resolved": "https://registry.npmjs.org/@napi-rs/cli/-/cli-3.7.0.tgz",
+      "integrity": "sha512-3d3+rmxlOIV/G1zPWeX4PCxuYnhcCQM2BvY9rtimC8RO0dFR9gtYP+Grov+WoduZtfWRj5N1XvytWeRxxCk5zw==",
       "dev": true,
       "license": "MIT",
       "dependencies": {
@@ -2954,7 +2954,7 @@
         "@octokit/rest": "^22.0.1",
         "clipanion": "^4.0.0-rc.4",
         "colorette": "^2.0.20",
-        "emnapi": "^1.7.1",
+        "emnapi": "^1.10.0",
         "es-toolkit": "^1.41.0",
         "js-yaml": "^4.1.0",
         "obug": "^2.0.0",

nodejs/package.json

@@ -11,7 +11,7 @@
     "ann"
   ],
   "private": false,
-  "version": "0.30.0-beta.0",
+  "version": "0.30.1-beta.2",
   "main": "dist/index.js",
   "exports": {
     ".": "./dist/index.js",
@@ -43,7 +43,7 @@
     "@aws-sdk/client-s3": "3.1003.0",
     "@biomejs/biome": "^1.7.3",
     "@jest/globals": "^29.7.0",
-    "@napi-rs/cli": "3.5.1",
+    "@napi-rs/cli": "3.7.0",
     "@types/axios": "^0.14.0",
     "@types/jest": "^29.1.2",
     "@types/node": "22.7.4",

nodejs/pnpm-lock.yaml

@@ -31,8 +31,8 @@ importers:
         specifier: ^29.7.0
         version: 29.7.0
       '@napi-rs/cli':
-        specifier: 3.5.1
-        version: 3.5.1(@emnapi/core@1.10.0)(@emnapi/runtime@1.10.0)(@types/node@22.7.4)
+        specifier: 3.7.0
+        version: 3.7.0(@emnapi/core@1.10.0)(@emnapi/runtime@1.10.0)(@types/node@22.7.4)
       '@types/axios':
         specifier: ^0.14.0
         version: 0.14.4
@@ -887,8 +887,8 @@ packages:
   '@jridgewell/trace-mapping@0.3.31':
     resolution: {integrity: sha512-zzNR+SdQSDJzc8joaeP8QQoCQr8NuYx2dIIytl1QeBEZHJ9uW6hebsrYgbz8hJwUQao3TWCMtmfV8Nu1twOLAw==}
 
-  '@napi-rs/cli@3.5.1':
-    resolution: {integrity: sha512-XBfLQRDcB3qhu6bazdMJsecWW55kR85l5/k0af9BIBELXQSsCFU0fzug7PX8eQp6vVdm7W/U3z6uP5WmITB2Gw==}
+  '@napi-rs/cli@3.7.0':
+    resolution: {integrity: sha512-3d3+rmxlOIV/G1zPWeX4PCxuYnhcCQM2BvY9rtimC8RO0dFR9gtYP+Grov+WoduZtfWRj5N1XvytWeRxxCk5zw==}
     engines: {node: '>= 16'}
     hasBin: true
     peerDependencies:
@@ -4582,7 +4582,7 @@ snapshots:
       '@jridgewell/resolve-uri': 3.1.2
       '@jridgewell/sourcemap-codec': 1.5.5
 
-  '@napi-rs/cli@3.5.1(@emnapi/core@1.10.0)(@emnapi/runtime@1.10.0)(@types/node@22.7.4)':
+  '@napi-rs/cli@3.7.0(@emnapi/core@1.10.0)(@emnapi/runtime@1.10.0)(@types/node@22.7.4)':
     dependencies:
       '@inquirer/prompts': 8.4.3(@types/node@22.7.4)
       '@napi-rs/cross-toolchain': 1.0.3(@emnapi/core@1.10.0)(@emnapi/runtime@1.10.0)

nodejs/src/lib.rs

@@ -24,15 +24,19 @@ mod util;
 #[napi(object)]
 #[derive(Debug)]
 pub struct ConnectionOptions {
-    /// (For LanceDB OSS only): The interval, in seconds, at which to check for
-    /// updates to the table from other processes. If None, then consistency is not
-    /// checked. For performance reasons, this is the default. For strong
-    /// consistency, set this to zero seconds. Then every read will check for
-    /// updates from other processes. As a compromise, you can set this to a
-    /// non-zero value for eventual consistency. If more than that interval
-    /// has passed since the last check, then the table will be checked for updates.
-    /// Note: this consistency only applies to read operations. Write operations are
+    /// The interval, in seconds, at which to check for updates to the table
+    /// from other processes. If None, then consistency is not checked. For
+    /// performance reasons, this is the default. For strong consistency, set
+    /// this to zero seconds. Then every read will check for updates from other
+    /// processes. As a compromise, you can set this to a non-zero value for
+    /// eventual consistency. If more than that interval has passed since the
+    /// last check, then the table will be checked for updates. Note: this
+    /// consistency only applies to read operations. Write operations are
     /// always consistent.
+    ///
+    /// Stronger consistency is not free. The smaller the interval, the more
+    /// often each read pays the cost of checking for updates against object
+    /// storage, raising per-read latency and cost.
     pub read_consistency_interval: Option<f64>,
     /// (For LanceDB OSS only): configuration for object storage.
     ///

nodejs/src/merge.rs

@@ -50,6 +50,20 @@ impl NativeMergeInsertBuilder {
         this
     }
 
+    #[napi]
+    pub fn use_lsm_write(&self, use_lsm_write: bool) -> Self {
+        let mut this = self.clone();
+        this.inner.use_lsm_write(use_lsm_write);
+        this
+    }
+
+    #[napi]
+    pub fn validate_single_shard(&self, validate_single_shard: bool) -> Self {
+        let mut this = self.clone();
+        this.inner.validate_single_shard(validate_single_shard);
+        this
+    }
+
     #[napi(catch_unwind)]
     pub async fn execute(&self, buf: Buffer) -> napi::Result<MergeResult> {
         let data = ipc_file_to_batches(buf.to_vec())

nodejs/src/table.rs

@@ -5,8 +5,9 @@ use std::collections::HashMap;
 
 use lancedb::ipc::{ipc_file_to_batches, ipc_file_to_schema};
 use lancedb::table::{
-    AddDataMode, ColumnAlteration as LanceColumnAlteration, Duration, NewColumnTransform,
-    OptimizeAction, OptimizeOptions, Table as LanceDbTable,
+    AddDataMode, ColumnAlteration as LanceColumnAlteration, Duration,
+    FieldMetadataUpdate as LanceFieldMetadataUpdate, NewColumnTransform, OptimizeAction,
+    OptimizeOptions, Ref, Table as LanceDbTable,
 };
 use napi::bindgen_prelude::*;
 use napi::threadsafe_function::{ThreadsafeFunction, ThreadsafeFunctionCallMode};
@@ -355,6 +356,23 @@ impl Table {
         Ok(res.into())
     }
 
+    #[napi(catch_unwind)]
+    pub async fn update_field_metadata(
+        &self,
+        updates: Vec<FieldMetadataUpdate>,
+    ) -> napi::Result<UpdateFieldMetadataResult> {
+        let updates = updates
+            .into_iter()
+            .map(LanceFieldMetadataUpdate::from)
+            .collect::<Vec<_>>();
+        let res = self
+            .inner_ref()?
+            .update_field_metadata(&updates)
+            .await
+            .default_error()?;
+        Ok(res.into())
+    }
+
     #[napi(catch_unwind)]
     pub async fn drop_columns(&self, columns: Vec<String>) -> napi::Result<DropColumnsResult> {
         let col_refs = columns.iter().map(String::as_str).collect::<Vec<_>>();
@@ -391,6 +409,11 @@ impl Table {
             .default_error()
     }
 
+    #[napi(catch_unwind)]
+    pub async fn close_lsm_writers(&self) -> napi::Result<()> {
+        self.inner_ref()?.close_lsm_writers().await.default_error()
+    }
+
     #[napi(catch_unwind)]
     pub async fn version(&self) -> napi::Result<i64> {
         self.inner_ref()?
@@ -455,6 +478,13 @@ impl Table {
         })
     }
 
+    #[napi(catch_unwind)]
+    pub async fn branches(&self) -> napi::Result<Branches> {
+        Ok(Branches {
+            inner: self.inner_ref()?.clone(),
+        })
+    }
+
     #[napi(catch_unwind)]
     pub async fn optimize(
         &self,
@@ -742,6 +772,29 @@ pub struct ColumnAlteration {
     pub nullable: Option<bool>,
 }
 
+/// A per-field metadata update, addressed by dot-path. Merges into the field's
+/// existing metadata by default; a `null` value deletes a key, and `replace`
+/// swaps the field's entire metadata map.
+#[napi(object)]
+pub struct FieldMetadataUpdate {
+    /// Dot-separated path to the field (e.g. "embedding" or "a.b.c").
+    pub path: String,
+    /// Metadata keys to set; a `null` value deletes that key.
+    pub metadata: HashMap<String, Option<String>>,
+    /// If true, replace the field's entire metadata map instead of merging.
+    pub replace: Option<bool>,
+}
+
+impl From<FieldMetadataUpdate> for LanceFieldMetadataUpdate {
+    fn from(js: FieldMetadataUpdate) -> Self {
+        Self {
+            path: js.path,
+            metadata: js.metadata,
+            replace: js.replace.unwrap_or(false),
+        }
+    }
+}
+
 impl TryFrom<ColumnAlteration> for LanceColumnAlteration {
     type Error = String;
     fn try_from(js: ColumnAlteration) -> std::result::Result<Self, Self::Error> {
@@ -792,9 +845,6 @@ pub struct IndexStatistics {
     pub distance_type: Option<String>,
     /// The number of parts this index is split into.
     pub num_indices: Option<u32>,
-    /// The KMeans loss value of the index,
-    /// it is only present for vector indices.
-    pub loss: Option<f64>,
 }
 impl From<lancedb::index::IndexStatistics> for IndexStatistics {
     fn from(value: lancedb::index::IndexStatistics) -> Self {
@@ -804,7 +854,6 @@ impl From<lancedb::index::IndexStatistics> for IndexStatistics {
             index_type: value.index_type.to_string(),
             distance_type: value.distance_type.map(|d| d.to_string()),
             num_indices: value.num_indices,
-            loss: value.loss,
         }
     }
 }
@@ -940,6 +989,7 @@ pub struct MergeResult {
     pub num_updated_rows: i64,
     pub num_deleted_rows: i64,
     pub num_attempts: i64,
+    pub num_rows: i64,
 }
 
 impl From<lancedb::table::MergeResult> for MergeResult {
@@ -950,6 +1000,7 @@ impl From<lancedb::table::MergeResult> for MergeResult {
             num_updated_rows: value.num_updated_rows as i64,
             num_deleted_rows: value.num_deleted_rows as i64,
             num_attempts: value.num_attempts as i64,
+            num_rows: value.num_rows as i64,
         }
     }
 }
@@ -980,6 +1031,19 @@ impl From<lancedb::table::AlterColumnsResult> for AlterColumnsResult {
     }
 }
 
+#[napi(object)]
+pub struct UpdateFieldMetadataResult {
+    pub version: i64,
+}
+
+impl From<lancedb::table::UpdateFieldMetadataResult> for UpdateFieldMetadataResult {
+    fn from(value: lancedb::table::UpdateFieldMetadataResult) -> Self {
+        Self {
+            version: value.version as i64,
+        }
+    }
+}
+
 #[napi(object)]
 pub struct DropColumnsResult {
     pub version: i64,
@@ -999,6 +1063,13 @@ pub struct TagContents {
     pub manifest_size: i64,
 }
 
+#[napi]
+pub struct BranchContents {
+    pub parent_branch: Option<String>,
+    pub parent_version: i64,
+    pub manifest_size: i64,
+}
+
 #[napi]
 pub struct Tags {
     inner: LanceDbTable,
@@ -1067,3 +1138,75 @@ impl Tags {
             .default_error()
     }
 }
+
+#[napi]
+pub struct Branches {
+    inner: LanceDbTable,
+}
+
+#[napi]
+impl Branches {
+    #[napi]
+    pub async fn list(&self) -> napi::Result<HashMap<String, BranchContents>> {
+        let branches = self.inner.list_branches().await.default_error()?;
+        let result = branches
+            .into_iter()
+            .map(|(k, v)| {
+                (
+                    k,
+                    BranchContents {
+                        parent_branch: v.parent_branch,
+                        parent_version: v.parent_version as i64,
+                        manifest_size: v.manifest_size as i64,
+                    },
+                )
+            })
+            .collect();
+        Ok(result)
+    }
+
+    #[napi]
+    pub async fn create(
+        &self,
+        name: String,
+        from_ref: Option<String>,
+        from_version: Option<i64>,
+    ) -> napi::Result<Table> {
+        let from_ref = from_ref.filter(|b| b != "main");
+        let from_version = from_version
+            .map(|v| {
+                u64::try_from(v).map_err(|_| {
+                    napi::Error::from_reason("from_version must be a non-negative integer")
+                })
+            })
+            .transpose()?;
+        let from = Ref::Version(from_ref, from_version);
+        let table = self
+            .inner
+            .create_branch(&name, from)
+            .await
+            .default_error()?;
+        Ok(Table::new(table))
+    }
+
+    #[napi]
+    pub async fn checkout(&self, name: String, version: Option<i64>) -> napi::Result<Table> {
+        let version = version
+            .map(|v| {
+                u64::try_from(v)
+                    .map_err(|_| napi::Error::from_reason("version must be a non-negative integer"))
+            })
+            .transpose()?;
+        let table = self
+            .inner
+            .checkout_branch(&name, version)
+            .await
+            .default_error()?;
+        Ok(Table::new(table))
+    }
+
+    #[napi]
+    pub async fn delete(&self, name: String) -> napi::Result<()> {
+        self.inner.delete_branch(&name).await.default_error()
+    }
+}

python/.bumpversion.toml

@@ -1,5 +1,5 @@
 [tool.bumpversion]
-current_version = "0.33.0-beta.1"
+current_version = "0.33.1-beta.2"
 parse = """(?x)
     (?P<major>0|[1-9]\\d*)\\.
     (?P<minor>0|[1-9]\\d*)\\.

python/Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "lancedb-python"
-version = "0.33.0-beta.1"
+version = "0.33.1-beta.2"
 publish = false
 edition.workspace = true
 description = "Python bindings for LanceDB"

python/python/lancedb/__init__.py

@@ -94,7 +94,6 @@ def connect(
     host_override: str, optional
         The override url for LanceDB Cloud.
     read_consistency_interval: timedelta, default None
-        (For LanceDB OSS only)
         The interval at which to check for updates to the table from other
         processes. If None, then consistency is not checked. For performance
         reasons, this is the default. For strong consistency, set this to
@@ -104,6 +103,10 @@ def connect(
         the last check, then the table will be checked for updates. Note: this
         consistency only applies to read operations. Write operations are
         always consistent.
+
+        Stronger consistency is not free. The smaller the interval, the more
+        often each read pays the cost of checking for updates against object
+        storage, raising per-read latency and cost.
     client_config: ClientConfig or dict, optional
         Configuration options for the LanceDB Cloud HTTP client. If a dict, then
         the keys are the attributes of the ClientConfig class. If None, then the
@@ -147,6 +150,13 @@ def connect(
     >>> db = lancedb.connect("s3://my-bucket/lancedb",
     ...                      storage_options={"aws_access_key_id": "***"})
 
+    For tests and temporary data, use an in-memory database:
+
+    >>> db = lancedb.connect("memory://")
+
+    In-memory databases are not persisted. Tables are dropped when the last
+    connection or table handle referencing them is closed.
+
     Connect to LanceDB cloud:
 
     >>> db = lancedb.connect("db://my_database", api_key="ldb_...",
@@ -210,6 +220,7 @@ def connect(
             request_thread_pool=request_thread_pool,
             client_config=client_config,
             storage_options=storage_options,
+            read_consistency_interval=read_consistency_interval,
             **kwargs,
         )
     _check_s3_bucket_with_dots(str(uri), storage_options)
@@ -304,6 +315,15 @@ def deserialize_conn(
             manifest_enabled=parsed.get("manifest_enabled", False),
             namespace_client_properties=parsed.get("namespace_client_properties"),
         )
+    elif connection_type == "remote":
+        return RemoteDBConnection(
+            parsed["db_url"],
+            parsed["api_key"],
+            parsed.get("region", "us-east-1"),
+            host_override=parsed.get("host_override"),
+            client_config=parsed.get("client_config"),
+            storage_options=storage_options,
+        )
     else:
         raise ValueError(f"Unknown connection_type: {connection_type}")
 
@@ -336,7 +356,6 @@ async def connect_async(
     host_override: str, optional
         The override url for LanceDB Cloud.
     read_consistency_interval: timedelta, default None
-        (For LanceDB OSS only)
         The interval at which to check for updates to the table from other
         processes. If None, then consistency is not checked. For performance
         reasons, this is the default. For strong consistency, set this to
@@ -346,6 +365,10 @@ async def connect_async(
         the last check, then the table will be checked for updates. Note: this
         consistency only applies to read operations. Write operations are
         always consistent.
+
+        Stronger consistency is not free. The smaller the interval, the more
+        often each read pays the cost of checking for updates against object
+        storage, raising per-read latency and cost.
     client_config: ClientConfig or dict, optional
         Configuration options for the LanceDB Cloud HTTP client. If a dict, then
         the keys are the attributes of the ClientConfig class. If None, then the
@@ -378,6 +401,8 @@ async def connect_async(
     ...     db = await lancedb.connect_async("s3://my-bucket/lancedb",
     ...                                      storage_options={
     ...                                          "aws_access_key_id": "***"})
+    ...     # For tests and temporary data, use an in-memory database
+    ...     db = await lancedb.connect_async("memory://")
     ...     # Connect to LanceDB cloud
     ...     db = await lancedb.connect_async("db://my_database", api_key="ldb_...",
     ...                                      client_config={

python/python/lancedb/_lancedb.pyi

@@ -208,6 +208,9 @@ class Table:
     async def alter_columns(
         self, columns: list[dict[str, Any]]
     ) -> AlterColumnsResult: ...
+    async def update_field_metadata(
+        self, updates: list[dict[str, Any]]
+    ) -> UpdateFieldMetadataResult: ...
     async def optimize(
         self,
         *,
@@ -220,8 +223,12 @@ class Table:
     async def set_unenforced_primary_key(self, columns: List[str]) -> None: ...
     async def set_lsm_write_spec(self, spec: LsmWriteSpec) -> None: ...
     async def unset_lsm_write_spec(self) -> None: ...
+    async def close_lsm_writers(self) -> None: ...
     @property
     def tags(self) -> Tags: ...
+    @property
+    def branches(self) -> Branches: ...
+    def current_branch(self) -> Optional[str]: ...
     def query(self) -> Query: ...
     def take_offsets(self, offsets: list[int]) -> TakeQuery: ...
     def take_row_ids(self, row_ids: list[int]) -> TakeQuery: ...
@@ -234,6 +241,17 @@ class Tags:
     async def delete(self, tag: str): ...
     async def update(self, tag: str, version: int): ...
 
+class Branches:
+    async def list(self) -> Dict[str, Any]: ...
+    async def create(
+        self,
+        name: str,
+        from_ref: Optional[str] = None,
+        from_version: Optional[int] = None,
+    ) -> Table: ...
+    async def checkout(self, name: str, version: Optional[int] = None) -> Table: ...
+    async def delete(self, name: str) -> None: ...
+
 class IndexConfig:
     name: str
     index_type: str
@@ -420,6 +438,7 @@ class MergeResult:
     num_inserted_rows: int
     num_deleted_rows: int
     num_attempts: int
+    num_rows: int
 
 class LsmWriteSpec:
     """Specification selecting Lance's MemWAL LSM-style write path for
@@ -458,6 +477,9 @@ class AddColumnsResult:
 class AlterColumnsResult:
     version: int
 
+class UpdateFieldMetadataResult:
+    version: int
+
 class DropColumnsResult:
     version: int
 

python/python/lancedb/background_loop.py

@@ -2,6 +2,7 @@
 # SPDX-FileCopyrightText: Copyright The LanceDB Authors
 
 import asyncio
+import concurrent.futures
 import os
 import threading
 import warnings
@@ -37,6 +38,24 @@ class BackgroundEventLoop:
 
 LOOP = BackgroundEventLoop()
 
+
+def _new_embedding_executor() -> concurrent.futures.ThreadPoolExecutor:
+    return concurrent.futures.ThreadPoolExecutor(thread_name_prefix="lancedb-embedding")
+
+
+# Embedding functions can block for a long time -- a heavy local model or an
+# HTTP request to a remote embeddings API. Running them on asyncio's default
+# executor lets them starve the unrelated blocking I/O that shares that pool,
+# so they get a dedicated one. See
+# https://github.com/lancedb/lancedb/issues/3310.
+_EMBEDDING_EXECUTOR = _new_embedding_executor()
+
+
+def embedding_executor() -> concurrent.futures.ThreadPoolExecutor:
+    """Return the executor dedicated to running blocking embedding calls."""
+    return _EMBEDDING_EXECUTOR
+
+
 _FORK_WARNED = False
 
 
@@ -47,6 +66,12 @@ def _reset_after_fork():
     # the new state. The Rust-side tokio runtime is reset analogously by a
     # pthread_atfork hook installed in the _lancedb extension.
     LOOP._start()
+    # The embedding executor's worker threads are dead in the child as well.
+    # Replace it with a fresh pool (threads are spawned lazily, so this is
+    # cheap); we don't shut down the old one, since joining its dead workers
+    # could hang.
+    global _EMBEDDING_EXECUTOR
+    _EMBEDDING_EXECUTOR = _new_embedding_executor()
     global _FORK_WARNED
     if not _FORK_WARNED:
         _FORK_WARNED = True

python/python/lancedb/db.py

@@ -416,6 +416,8 @@ class DBConnection(EnforceOverrides):
         namespace_path: Optional[List[str]] = None,
         storage_options: Optional[Dict[str, str]] = None,
         index_cache_size: Optional[int] = None,
+        branch: Optional[str] = None,
+        version: Optional[int] = None,
     ) -> Table:
         """Open a Lance Table in the database.
 
@@ -444,6 +446,14 @@ class DBConnection(EnforceOverrides):
             connection will be inherited by the table, but can be overridden here.
             See available options at
             <https://docs.lancedb.com/storage/>
+        branch: str, optional
+            If provided, open a handle scoped to this branch instead of the
+            default branch. Reads and writes operate in the branch's context.
+        version: int, optional
+            If provided, open the table pinned to this version, producing a
+            read-only handle. Composes with ``branch``: when both are given,
+            opens that branch at the version; otherwise opens ``main`` at the
+            version. Call ``checkout_latest`` to return to a writable state.
 
         Returns
         -------
@@ -958,6 +968,8 @@ class LanceDBConnection(DBConnection):
         namespace_path: Optional[List[str]] = None,
         storage_options: Optional[Dict[str, str]] = None,
         index_cache_size: Optional[int] = None,
+        branch: Optional[str] = None,
+        version: Optional[int] = None,
     ) -> LanceTable:
         """Open a table in the database.
 
@@ -968,6 +980,14 @@ class LanceDBConnection(DBConnection):
         namespace_path: List[str], optional
             The namespace to open the table from.  When non-empty, the
             table is resolved through the directory namespace client.
+        branch: str, optional
+            If provided, open a handle scoped to this branch instead of the
+            default branch. Reads and writes operate in the branch's context.
+        version: int, optional
+            If provided, open the table pinned to this version, producing a
+            read-only handle. Composes with ``branch``: when both are given,
+            opens that branch at the version; otherwise opens ``main`` at the
+            version. Call ``checkout_latest`` to return to a writable state.
 
         Returns
         -------
@@ -987,20 +1007,26 @@ class LanceDBConnection(DBConnection):
             )
 
         if namespace_path:
-            return self._namespace_conn().open_table(
+            tbl = self._namespace_conn().open_table(
+                name,
+                namespace_path=namespace_path,
+                storage_options=storage_options,
+                index_cache_size=index_cache_size,
+            )
+        else:
+            tbl = LanceTable.open(
+                self,
                 name,
                 namespace_path=namespace_path,
                 storage_options=storage_options,
                 index_cache_size=index_cache_size,
             )
 
-        return LanceTable.open(
-            self,
-            name,
-            namespace_path=namespace_path,
-            storage_options=storage_options,
-            index_cache_size=index_cache_size,
-        )
+        if branch is not None:
+            tbl = tbl.branches.checkout(branch, version)
+        elif version is not None:
+            tbl.checkout(version)
+        return tbl
 
     def clone_table(
         self,
@@ -1641,6 +1667,8 @@ class AsyncConnection(object):
         location: Optional[str] = None,
         namespace_client: Optional[Any] = None,
         managed_versioning: Optional[bool] = None,
+        branch: Optional[str] = None,
+        version: Optional[int] = None,
     ) -> AsyncTable:
         """Open a Lance Table in the database.
 
@@ -1676,6 +1704,14 @@ class AsyncConnection(object):
         managed_versioning: bool, optional
             Whether managed versioning is enabled for this table. If provided,
             avoids a redundant describe_table call when namespace_client is set.
+        branch: str, optional
+            If provided, open a handle scoped to this branch instead of the
+            default branch. Reads and writes operate in the branch's context.
+        version: int, optional
+            If provided, open the table pinned to this version, producing a
+            read-only handle. Composes with ``branch``: when both are given,
+            opens that branch at the version; otherwise opens ``main`` at the
+            version. Call ``checkout_latest`` to return to a writable state.
 
         Returns
         -------
@@ -1692,7 +1728,14 @@ class AsyncConnection(object):
             namespace_client=namespace_client,
             managed_versioning=managed_versioning,
         )
-        return AsyncTable(table)
+        tbl = AsyncTable(table)
+        # "main" is the default branch, so treat it as no branch: remote rejects
+        # every branch checkout (even "main"), and the version still applies.
+        if branch is not None and branch != "main":
+            tbl = await tbl.branches.checkout(branch, version)
+        elif version is not None:
+            await tbl.checkout(version)
+        return tbl
 
     async def clone_table(
         self,

python/python/lancedb/index.py

@@ -281,6 +281,9 @@ class HnswPq:
     m: int = 20
     ef_construction: int = 300
     target_partition_size: Optional[int] = None
+    # Name of the accelerator (e.g. "cuda") to use for IVF training. When set,
+    # create_index() dispatches to pylance to build the index on the accelerator.
+    accelerator: Optional[str] = None
 
 
 @dataclass
@@ -386,6 +389,9 @@ class HnswSq:
     m: int = 20
     ef_construction: int = 300
     target_partition_size: Optional[int] = None
+    # Name of the accelerator (e.g. "cuda") to use for IVF training. When set,
+    # create_index() dispatches to pylance to build the index on the accelerator.
+    accelerator: Optional[str] = None
 
 
 @dataclass
@@ -579,6 +585,9 @@ class IvfFlat:
     max_iterations: int = 50
     sample_rate: int = 256
     target_partition_size: Optional[int] = None
+    # Name of the accelerator (e.g. "cuda") to use for IVF training. When set,
+    # create_index() dispatches to pylance to build the index on the accelerator.
+    accelerator: Optional[str] = None
 
 
 @dataclass
@@ -609,6 +618,9 @@ class IvfSq:
     max_iterations: int = 50
     sample_rate: int = 256
     target_partition_size: Optional[int] = None
+    # Name of the accelerator (e.g. "cuda") to use for IVF training. When set,
+    # create_index() dispatches to pylance to build the index on the accelerator.
+    accelerator: Optional[str] = None
 
 
 @dataclass
@@ -739,6 +751,9 @@ class IvfPq:
     max_iterations: int = 50
     sample_rate: int = 256
     target_partition_size: Optional[int] = None
+    # Name of the accelerator (e.g. "cuda") to use for IVF training. When set,
+    # create_index() dispatches to pylance to build the index on the accelerator.
+    accelerator: Optional[str] = None
 
 
 @dataclass
@@ -792,6 +807,9 @@ class IvfRq:
     max_iterations: int = 50
     sample_rate: int = 256
     target_partition_size: Optional[int] = None
+    # Name of the accelerator (e.g. "cuda") to use for IVF training. When set,
+    # create_index() dispatches to pylance to build the index on the accelerator.
+    accelerator: Optional[str] = None
 
 
 __all__ = [

python/python/lancedb/merge.py

@@ -34,6 +34,8 @@ class LanceMergeInsertBuilder(object):
         self._when_not_matched_by_source_condition = None
         self._timeout = None
         self._use_index = True
+        self._use_lsm_write = None
+        self._validate_single_shard = None
 
     def when_matched_update_all(
         self, *, where: Optional[str] = None
@@ -96,6 +98,46 @@ class LanceMergeInsertBuilder(object):
         self._use_index = use_index
         return self
 
+    def use_lsm_write(self, use_lsm_write: bool) -> LanceMergeInsertBuilder:
+        """
+        Controls whether the merge uses the MemWAL LSM write path.
+
+        By default (unset), a `merge_insert` on a table with an LSM write spec
+        is routed through Lance's MemWAL shard writer, and a table without one
+        uses the standard path. Pass `False` to force the standard path even
+        when a spec is set. Pass `True` to require a spec — `merge_insert`
+        raises an error if none is installed.
+
+        Parameters
+        ----------
+        use_lsm_write: bool
+            Whether to use the LSM write path.
+        """
+        self._use_lsm_write = use_lsm_write
+        return self
+
+    def validate_single_shard(
+        self, validate_single_shard: bool
+    ) -> LanceMergeInsertBuilder:
+        """
+        Controls how an LSM merge checks that its input targets a single shard.
+
+        When a table has an LSM write spec, every row in a `merge_insert` call
+        must route to the same shard. When `True` (the default), every row is
+        inspected to verify this. When `False`, only the first row is inspected
+        and the shard it routes to is used for the whole input — a faster path
+        for callers that have already pre-sharded their input.
+
+        Has no effect on tables without an LSM write spec.
+
+        Parameters
+        ----------
+        validate_single_shard: bool
+            Whether to check every row routes to one shard. Defaults to `True`.
+        """
+        self._validate_single_shard = validate_single_shard
+        return self
+
     def execute(
         self,
         new_data: DATA,

python/python/lancedb/namespace.py

@@ -144,7 +144,12 @@ def _query_to_namespace_request(
     if query.postfilter is not None:
         prefilter = not query.postfilter
 
-    k = query.limit if query.limit is not None else 10
+    if query.limit is not None:
+        k = query.limit
+    elif query.vector is None and query.full_text_query is None:
+        k = sys.maxsize
+    else:
+        k = 10
 
     # Build request kwargs, only including non-None values for optional fields
     # that Pydantic doesn't accept as None
@@ -544,6 +549,8 @@ class LanceNamespaceDBConnection(DBConnection):
         namespace_path: Optional[List[str]] = None,
         storage_options: Optional[Dict[str, str]] = None,
         index_cache_size: Optional[int] = None,
+        branch: Optional[str] = None,
+        version: Optional[int] = None,
     ) -> Table:
         if namespace_path is None:
             namespace_path = []
@@ -562,7 +569,7 @@ class LanceNamespaceDBConnection(DBConnection):
                 raise TableNotFoundError(f"Table not found: {'$'.join(table_id)}")
             raise
 
-        return LanceTable(
+        tbl = LanceTable(
             self,
             name,
             namespace_path=namespace_path,
@@ -570,6 +577,11 @@ class LanceNamespaceDBConnection(DBConnection):
             pushdown_operations=self._namespace_client_pushdown_operations,
             _async=async_table,
         )
+        if branch is not None:
+            tbl = tbl.branches.checkout(branch, version)
+        elif version is not None:
+            tbl.checkout(version)
+        return tbl
 
     @override
     def drop_table(self, name: str, namespace_path: Optional[List[str]] = None):
@@ -954,7 +966,7 @@ class AsyncLanceNamespaceDBConnection:
         if mode.lower() not in ["create", "overwrite"]:
             raise ValueError("mode must be either 'create' or 'overwrite'")
         validate_table_name(name)
-        return await self._inner.create_table(
+        table = await self._inner.create_table(
             name,
             data,
             schema=schema,
@@ -966,6 +978,11 @@ class AsyncLanceNamespaceDBConnection:
             embedding_functions=embedding_functions,
             storage_options=storage_options,
         )
+        return table._set_namespace_context(
+            namespace_path=namespace_path,
+            namespace_client=self._namespace_client,
+            pushdown_operations=self._namespace_client_pushdown_operations,
+        )
 
     async def open_table(
         self,
@@ -974,12 +991,14 @@ class AsyncLanceNamespaceDBConnection:
         namespace_path: Optional[List[str]] = None,
         storage_options: Optional[Dict[str, str]] = None,
         index_cache_size: Optional[int] = None,
+        branch: Optional[str] = None,
+        version: Optional[int] = None,
     ) -> AsyncTable:
         """Open an existing table from the namespace."""
         if namespace_path is None:
             namespace_path = []
         try:
-            return await self._inner.open_table(
+            table = await self._inner.open_table(
                 name,
                 namespace_path=namespace_path,
                 storage_options=storage_options,
@@ -990,6 +1009,17 @@ class AsyncLanceNamespaceDBConnection:
                 table_id = namespace_path + [name]
                 raise TableNotFoundError(f"Table not found: {'$'.join(table_id)}")
             raise
+        # "main" is the default branch, so treat it as no branch (mirrors the
+        # sync remote path); the version still applies.
+        if branch is not None and branch != "main":
+            table = await table.branches.checkout(branch, version)
+        elif version is not None:
+            await table.checkout(version)
+        return table._set_namespace_context(
+            namespace_path=namespace_path,
+            namespace_client=self._namespace_client,
+            pushdown_operations=self._namespace_client_pushdown_operations,
+        )
 
     async def drop_table(self, name: str, namespace_path: Optional[List[str]] = None):
         """Drop a table from the namespace."""

python/python/lancedb/permutation.py

@@ -3,12 +3,13 @@
 
 import copy
 import json
+import os
 
 from deprecation import deprecated
 import pyarrow as pa
 
 from ._lancedb import async_permutation_builder, PermutationReader
-from .table import LanceTable
+from .table import LanceTable, Table
 from .background_loop import LOOP
 from .util import batch_to_tensor, batch_to_tensor_rows
 from typing import Any, Callable, Iterator, Literal, Optional, TYPE_CHECKING, Union
@@ -354,6 +355,49 @@ class Transforms:
 DEFAULT_BATCH_SIZE = 100
 
 
+def _table_to_pickle_state(table: Table) -> dict[str, Any]:
+    from .remote.table import RemoteTable
+
+    if isinstance(table, RemoteTable):
+        return {
+            "kind": "remote",
+            "table": table,
+        }
+
+    if not isinstance(table, LanceTable):
+        raise ValueError(f"Cannot pickle table of type {type(table)!r}")
+
+    base_uri = table._conn.uri
+    if base_uri.startswith("memory://"):
+        return {
+            "kind": "memory",
+            "name": table.name,
+            "data": table.to_arrow(),
+        }
+
+    return {
+        "kind": "local",
+        "name": table.name,
+        "uri": base_uri,
+        "namespace": table._namespace_path,
+        "storage_options": table._conn.storage_options,
+    }
+
+
+def _table_from_pickle_state(state: dict[str, Any]) -> Table:
+    from . import connect
+
+    kind = state["kind"]
+    if kind == "remote":
+        return state["table"]
+    if kind == "memory":
+        return connect("memory://").create_table(state["name"], state["data"])
+    if kind == "local":
+        db = connect(state["uri"], storage_options=state["storage_options"])
+        return db.open_table(state["name"], namespace_path=state["namespace"] or None)
+    raise ValueError(f"Unknown table pickle state kind: {kind}")
+
+
 class Permutation:
     """
     A Permutation is a view of a dataset that can be used as input to model training
@@ -369,15 +413,15 @@ class Permutation:
 
     def __init__(
         self,
-        base_table: LanceTable,
-        permutation_table: Optional[LanceTable],
+        base_table: Table,
+        permutation_table: Optional[Table],
         split: int,
         selection: dict[str, str],
         batch_size: int,
         transform_fn: Callable[pa.RecordBatch, Any],
         offset: Optional[int] = None,
         limit: Optional[int] = None,
-        connection_factory: Optional[Callable[[str], LanceTable]] = None,
+        connection_factory: Optional[Callable[[str], Table]] = None,
         _reader: Optional[PermutationReader] = None,
     ):
         """
@@ -397,6 +441,7 @@ class Permutation:
         if _reader is None:
             _reader = LOOP.run(self._build_reader())
         self.reader: PermutationReader = _reader
+        self._pid = os.getpid()
 
     async def _build_reader(self) -> PermutationReader:
         reader = await PermutationReader.from_tables(
@@ -428,29 +473,25 @@ class Permutation:
         return new
 
     def with_connection_factory(
-        self, connection_factory: Callable[[str], LanceTable]
+        self, connection_factory: Callable[[str], Table]
     ) -> "Permutation":
         """
         Creates a new permutation that will use ``connection_factory`` to reopen
         the base table when this permutation is unpickled in a worker process.
 
-        The factory is a callable that takes a single argument — the base table
-        name — and returns a [LanceTable]. It must be picklable; the worker
+        The factory is a callable that takes a single argument, the base table
+        name, and returns a LanceDB table. It must be picklable; the worker
         will pickle it via standard ``pickle`` and call it to recover the base
         table. Picklable callables in practice means top-level (module-level)
         functions, ``functools.partial`` of such functions, or instances of
         picklable classes implementing ``__call__``. Lambdas and closures over
         local variables don't pickle with the default protocol.
 
-        Setting a factory is necessary when the URI alone is not enough to
-        re-open the connection — most importantly for LanceDB Cloud (``db://``)
-        connections, where ``api_key`` and ``region`` aren't recoverable from
-        the connection object after construction.
-
-        For local file or cloud-storage paths the factory is optional: if not
-        set, ``__getstate__`` falls back to capturing
-        ``(uri, storage_options, namespace_path)`` and re-opening via
-        ``lancedb.connect(uri, storage_options=...)``.
+        A factory is optional for normal local and remote LanceDB connections:
+        if not set, ``__getstate__`` captures the table's own picklable reopen
+        state. Use a factory when that default state is not enough, for example
+        when credentials should be loaded from the worker environment instead
+        of being embedded in the pickle.
 
         Examples
         --------
@@ -508,7 +549,7 @@ class Permutation:
         return new
 
     @classmethod
-    def identity(cls, table: LanceTable) -> "Permutation":
+    def identity(cls, table: Table) -> "Permutation":
         """
         Creates an identity permutation for the given table.
         """
@@ -517,8 +558,8 @@ class Permutation:
     @classmethod
     def from_tables(
         cls,
-        base_table: LanceTable,
-        permutation_table: Optional[LanceTable] = None,
+        base_table: Table,
+        permutation_table: Optional[Table] = None,
         split: Optional[Union[str, int]] = None,
     ) -> "Permutation":
         """
@@ -594,11 +635,10 @@ class Permutation:
 
         The base table is captured either via a user-supplied
         ``connection_factory`` (see [with_connection_factory]) or, as a
-        fallback, by introspecting ``(uri, storage_options, namespace_path)``
-        on the connection. The permutation table — always an in-memory
-        LanceDB table — is captured as a pyarrow Table (which pickles via
-        Arrow IPC natively). The reader is dropped from the wire format;
-        ``__setstate__`` rebuilds it from the restored tables.
+        fallback, by the table's own picklable reopen state. The permutation
+        table is captured as a pyarrow Table (which pickles via Arrow IPC
+        natively). The reader is dropped from the wire format and rebuilt
+        lazily on first use.
         """
         permutation_data: Optional[pa.Table] = None
         if self.permutation_table is not None:
@@ -622,39 +662,9 @@ class Permutation:
             # namespace from the existing connection.
             return common
 
-        # URI-introspection fallback: only viable for native (OSS) connections
-        # where (uri, storage_options) is enough to reopen. Remote / cloud
-        # connections don't expose recoverable api_key / region — those users
-        # must call with_connection_factory().
-        try:
-            base_uri = self.base_table._conn.uri
-            storage_options = self.base_table._conn.storage_options
-        except AttributeError as e:
-            raise ValueError(
-                "Cannot pickle this Permutation: the base table's connection "
-                "does not expose a uri/storage_options, which usually means it "
-                "is a remote (LanceDB Cloud) connection. Call "
-                "Permutation.with_connection_factory(...) first to provide a "
-                "picklable callable that re-opens the base table from a worker "
-                "process."
-            ) from e
-
-        if base_uri.startswith("memory://"):
-            # In-memory base tables don't exist in any worker process by
-            # default, so dump the entire base table into the pickle. This
-            # can be expensive for large datasets — users with large
-            # in-memory base tables should either persist them or set a
-            # connection_factory.
-            return {
-                **common,
-                "base_table_data": self.base_table.to_arrow(),
-            }
-
         return {
             **common,
-            "base_table_uri": base_uri,
-            "base_table_namespace": self.base_table._namespace_path,
-            "base_table_storage_options": storage_options,
+            "base_table_state": _table_to_pickle_state(self.base_table),
         }
 
     def __setstate__(self, state: dict[str, Any]) -> None:
@@ -663,6 +673,8 @@ class Permutation:
         connection_factory = state["connection_factory"]
         if connection_factory is not None:
             base_table = connection_factory(state["base_table_name"])
+        elif "base_table_state" in state:
+            base_table = _table_from_pickle_state(state["base_table_state"])
         elif "base_table_data" in state:
             # In-memory base table inlined into the pickle; rebuild the same
             # way we rebuild the in-memory permutation table.
@@ -680,7 +692,7 @@ class Permutation:
                 namespace_path=state["base_table_namespace"] or None,
             )
 
-        permutation_table: Optional[LanceTable] = None
+        permutation_table: Optional[Table] = None
         if state["permutation_data"] is not None:
             mem_db = connect("memory://")
             permutation_table = mem_db.create_table(
@@ -696,10 +708,28 @@ class Permutation:
         self.offset = state["offset"]
         self.limit = state["limit"]
         self.connection_factory = connection_factory
+        self.reader = None
+        self._pid = None
+
+    def _ensure_open(self) -> None:
+        pid = os.getpid()
+        if self.reader is not None and getattr(self, "_pid", None) == pid:
+            return
+        # The reader owns Rust-side table handles. Rebuild it after unpickle or
+        # fork even though the Python table wrappers reopen themselves.
+        if hasattr(self.base_table, "_ensure_open"):
+            self.base_table._ensure_open()
+        if self.permutation_table is not None and hasattr(
+            self.permutation_table, "_ensure_open"
+        ):
+            self.permutation_table._ensure_open()
         self.reader = LOOP.run(self._build_reader())
+        self._pid = pid
 
     @property
     def schema(self) -> pa.Schema:
+        self._ensure_open()
+
         async def do_output_schema():
             return await self.reader.output_schema(self.selection)
 
@@ -717,6 +747,7 @@ class Permutation:
         """
         The number of rows in the permutation
         """
+        self._ensure_open()
         return self.reader.count_rows()
 
     @property
@@ -875,6 +906,7 @@ class Permutation:
         If skip_last_batch is True, the last batch will be skipped if it is not a
         multiple of batch_size.
         """
+        self._ensure_open()
 
         async def get_iter():
             return await self.reader.read(self.selection, batch_size=batch_size)
@@ -976,6 +1008,7 @@ class Permutation:
         so `with_format` and `with_transform` affect this method in the same way
         they affect iteration.
         """
+        self._ensure_open()
 
         async def do_take_offsets():
             return await self.reader.take_offsets(offsets, selection=self.selection)
@@ -1011,9 +1044,11 @@ class Permutation:
         """
         Skip the first `skip` rows of the permutation
         """
+        self._ensure_open()
         new = copy.copy(self)
         new.offset = skip
         new.reader = LOOP.run(new._build_reader())
+        new._pid = os.getpid()
         return new
 
     @deprecated(details="Use with_take instead")
@@ -1032,9 +1067,11 @@ class Permutation:
         """
         Limit the permutation to `limit` rows (following any `skip`)
         """
+        self._ensure_open()
         new = copy.copy(self)
         new.limit = limit
         new.reader = LOOP.run(new._build_reader())
+        new._pid = os.getpid()
         return new
 
     @deprecated(details="Use with_repeat instead")

python/python/lancedb/query.py

@@ -41,6 +41,14 @@ from .rerankers.rrf import RRFReranker
 from .rerankers.util import check_reranker_result
 from .util import flatten_columns
 
+BlobMode = Literal["lazy", "bytes", "descriptions"]
+
+_BLOB_MODE_TO_HANDLING = {
+    "lazy": "blobs_descriptions",
+    "bytes": "all_binary",
+    "descriptions": "blobs_descriptions",
+}
+
 if TYPE_CHECKING:
     import sys
 
@@ -55,7 +63,7 @@ if TYPE_CHECKING:
     from ._lancedb import VectorQuery as LanceVectorQuery
     from .common import VEC
     from .pydantic import LanceModel
-    from .table import Table
+    from .table import AsyncTable, Table
 
     if sys.version_info >= (3, 11):
         from typing import Self
@@ -65,6 +73,179 @@ if TYPE_CHECKING:
 T = TypeVar("T", bound="LanceModel")
 
 
+def _validate_blob_mode(blob_mode: BlobMode) -> None:
+    if blob_mode not in _BLOB_MODE_TO_HANDLING:
+        modes = ", ".join(repr(mode) for mode in _BLOB_MODE_TO_HANDLING)
+        raise ValueError(f"blob_mode must be one of {modes}, got {blob_mode!r}")
+
+
+def _field_is_blob(field: pa.Field) -> bool:
+    metadata = field.metadata or {}
+    return metadata.get(b"lance-encoding:blob") == b"true" or (
+        metadata.get("lance-encoding:blob") == "true"
+    )
+
+
+def _schema_has_blob_field(schema: pa.Schema) -> bool:
+    return any(_field_is_blob(field) for field in schema)
+
+
+def _blob_mode_requires_native_pandas(blob_mode: BlobMode, schema: pa.Schema) -> bool:
+    return blob_mode in _BLOB_MODE_TO_HANDLING and _schema_has_blob_field(schema)
+
+
+def _unsupported_blob_pandas_error(reason: str) -> RuntimeError:
+    return RuntimeError(
+        "blob columns require Lance native scanner conversion for query "
+        f"to_pandas(), but {reason}. Use a plain scan query or remove blob "
+        "columns from the projection."
+    )
+
+
+def _query_is_plain_scan(query: Query) -> bool:
+    return (
+        query.vector is None
+        and query.full_text_query is None
+        and not query.postfilter
+        and not query.order_by
+    )
+
+
+def _filter_to_sql(filter: Optional[Union[str, Expr]]) -> Optional[str]:
+    if filter is None:
+        return None
+    if isinstance(filter, Expr):
+        return filter.to_sql()
+    return filter
+
+
+def _projection_to_scanner_kwargs(
+    columns: Optional[
+        Union[
+            List[str], List[Tuple[str, Union[str, Expr]]], Dict[str, Union[str, Expr]]
+        ]
+    ],
+) -> Dict[str, Any]:
+    if columns is None:
+        return {}
+    if isinstance(columns, list):
+        if all(isinstance(column, str) for column in columns):
+            return {"columns": columns}
+        if all(isinstance(column, tuple) and len(column) == 2 for column in columns):
+            return {
+                "columns": {
+                    name: expr.to_sql() if isinstance(expr, Expr) else expr
+                    for name, expr in columns
+                }
+            }
+        # Let Lance raise the detailed projection validation error.
+        return {"columns": columns}
+
+    projection = {}
+    for name, expr in columns.items():
+        if isinstance(expr, Expr):
+            expr = expr.to_sql()
+        projection[name] = expr
+    return {"columns": projection}
+
+
+def _scanner_kwargs_for_query(
+    query: Query, blob_mode: BlobMode, dataset: Optional[Any] = None
+) -> Dict[str, Any]:
+    fragments = _scanner_fragments_for_query(query, dataset)
+    kwargs = {
+        **_projection_to_scanner_kwargs(query.columns),
+        "filter": _filter_to_sql(query.filter),
+        "limit": query.limit,
+        "offset": query.offset,
+        "with_row_id": query.with_row_id,
+        "with_row_address": query.with_row_address,
+        "fast_search": query.fast_search,
+        "blob_handling": _BLOB_MODE_TO_HANDLING[blob_mode],
+        "fragments": fragments,
+    }
+    return {key: value for key, value in kwargs.items() if value is not None}
+
+
+def _scanner_fragments_for_query(query: Query, dataset: Optional[Any]) -> Optional[Any]:
+    if query.fragments is not None and query.fragment_ids is not None:
+        raise ValueError("fragments and fragment_ids cannot both be set")
+    if query.fragments is not None:
+        return query.fragments
+    if query.fragment_ids is None:
+        return None
+    if dataset is None:
+        raise ValueError("fragment_ids require a Lance dataset")
+
+    requested = set(query.fragment_ids)
+    fragments = [
+        fragment
+        for fragment in dataset.get_fragments()
+        if fragment.fragment_id in requested
+    ]
+    found = {fragment.fragment_id for fragment in fragments}
+    missing = requested - found
+    if missing:
+        missing_ids = ", ".join(str(fragment_id) for fragment_id in sorted(missing))
+        raise ValueError(f"fragment_ids not found in dataset: {missing_ids}")
+    return fragments
+
+
+def _ensure_lazy_blob_frame(
+    df: "pd.DataFrame", schema: pa.Schema, blob_mode: BlobMode
+) -> "pd.DataFrame":
+    if blob_mode != "lazy" or not _schema_has_blob_field(schema) or len(df) == 0:
+        return df
+
+    for field in schema:
+        if not _field_is_blob(field) or field.name not in df.columns:
+            continue
+        value = df[field.name].iloc[0]
+        if value is not None and not hasattr(value, "readall"):
+            raise _unsupported_blob_pandas_error(
+                "the Lance scanner did not return lazy blob files"
+            )
+    return df
+
+
+def _scanner_to_table(scanner: Any) -> pa.Table:
+    if hasattr(scanner, "to_pyarrow"):
+        reader = scanner.to_pyarrow()
+        return reader.read_all()
+    if hasattr(scanner, "to_table"):
+        return scanner.to_table()
+    reader = scanner.to_reader()
+    return reader.read_all()
+
+
+def _scanner_to_pandas(scanner: Any, blob_mode: BlobMode, **kwargs) -> "pd.DataFrame":
+    schema = getattr(scanner, "projected_schema", None)
+    if schema is None:
+        schema = getattr(scanner, "schema", None)
+    if schema is None:
+        schema = getattr(scanner, "dataset_schema", None)
+    if callable(schema):
+        schema = schema()
+    if hasattr(scanner, "to_pandas"):
+        try:
+            df = scanner.to_pandas(blob_mode=blob_mode, **kwargs)
+        except TypeError as err:
+            message = str(err)
+            if "blob_mode" not in message and "unexpected keyword" not in message:
+                raise
+            df = scanner.to_pandas(**kwargs)
+        if schema is not None:
+            return _ensure_lazy_blob_frame(df, schema, blob_mode)
+        return df
+
+    tbl = _scanner_to_table(scanner)
+    if blob_mode == "lazy" and _schema_has_blob_field(tbl.schema):
+        raise _unsupported_blob_pandas_error(
+            "the Lance scanner does not expose to_pandas"
+        )
+    return tbl.to_pandas(**kwargs)
+
+
 # Pydantic validation function for vector queries
 def ensure_vector_query(
     val: Any,
@@ -499,6 +680,13 @@ class Query(pydantic.BaseModel):
     # if true, include the row id in the results
     with_row_id: Optional[bool] = None
 
+    # if true, include the row address in the results
+    with_row_address: Optional[bool] = None
+
+    # Lance fragments or fragment ids to scan on scanner-backed plain queries
+    fragments: Optional[Any] = None
+    fragment_ids: Optional[List[int]] = None
+
     # offset to start fetching results from
     offset: Optional[int] = None
 
@@ -691,6 +879,9 @@ class LanceQueryBuilder(ABC):
         self._where = None
         self._postfilter = None
         self._with_row_id = None
+        self._with_row_address = None
+        self._fragments = None
+        self._fragment_ids = None
         self._vector = None
         self._text = None
         self._ef = None
@@ -718,6 +909,7 @@ class LanceQueryBuilder(ABC):
         self,
         flatten: Optional[Union[int, bool]] = None,
         *,
+        blob_mode: BlobMode = "lazy",
         timeout: Optional[timedelta] = None,
         **kwargs,
     ) -> "pd.DataFrame":
@@ -737,11 +929,41 @@ class LanceQueryBuilder(ABC):
         timeout: Optional[timedelta]
             The maximum time to wait for the query to complete.
             If None, wait indefinitely.
+        blob_mode: str, default "lazy"
+            Controls how blob columns are returned for plain scan queries.
+            Vector, FTS, hybrid, and other non-native query shapes keep the
+            existing Arrow conversion path and only support blob descriptions.
         **kwargs
             Forwarded to pyarrow.Table.to_pandas after query execution and
             optional flattening.
         """
+        _validate_blob_mode(blob_mode)
+        output_schema = getattr(self, "output_schema", None)
+        if output_schema is not None:
+            schema = output_schema()
+            if _blob_mode_requires_native_pandas(blob_mode, schema):
+                native_error = None
+                if (flatten is None or blob_mode == "descriptions") and timeout is None:
+                    try:
+                        df = self._plain_scan_to_pandas(
+                            blob_mode, flatten=flatten, **kwargs
+                        )
+                        if df is not None:
+                            return df
+                    except Exception as err:
+                        native_error = err
+                reason = (
+                    "this query shape cannot use Lance native pandas conversion"
+                    if native_error is None
+                    else str(native_error)
+                )
+                raise _unsupported_blob_pandas_error(reason) from native_error
+
         tbl = flatten_columns(self.to_arrow(timeout=timeout), flatten)
+        if _blob_mode_requires_native_pandas(blob_mode, tbl.schema):
+            raise _unsupported_blob_pandas_error(
+                "this query shape cannot use Lance native pandas conversion"
+            )
         return tbl.to_pandas(**kwargs)
 
     @abstractmethod
@@ -947,6 +1169,32 @@ class LanceQueryBuilder(ABC):
         self._with_row_id = with_row_id
         return self
 
+    def with_row_address(self, with_row_address: bool = True) -> Self:
+        """Set whether to return row addresses.
+
+        Parameters
+        ----------
+        with_row_address: bool, default True
+            If True, return the _rowaddr column in the results.
+
+        Returns
+        -------
+        LanceQueryBuilder
+            The LanceQueryBuilder object.
+        """
+        self._with_row_address = with_row_address
+        return self
+
+    def with_fragments(self, fragments: Any) -> Self:
+        """Set the Lance fragments to scan for plain scanner-backed queries."""
+        self._fragments = fragments
+        return self
+
+    def fragment_ids(self, fragment_ids: List[int]) -> Self:
+        """Set the Lance fragment ids to scan for plain scanner-backed queries."""
+        self._fragment_ids = fragment_ids
+        return self
+
     def explain_plan(self, verbose: Optional[bool] = False) -> str:
         """Return the execution plan for this query.
 
@@ -1086,6 +1334,25 @@ class LanceQueryBuilder(ABC):
         """
         raise NotImplementedError
 
+    def _plain_scan_to_pandas(
+        self,
+        blob_mode: BlobMode,
+        flatten: Optional[Union[int, bool]] = None,
+        **kwargs,
+    ) -> Optional["pd.DataFrame"]:
+        query = self.to_query_object()
+        if not _query_is_plain_scan(query):
+            return None
+
+        dataset = self._table.to_lance()
+        scanner = dataset.scanner(
+            **_scanner_kwargs_for_query(query, blob_mode, dataset)
+        )
+        if flatten is not None:
+            tbl = flatten_columns(_scanner_to_table(scanner), flatten)
+            return tbl.to_pandas(**kwargs)
+        return _scanner_to_pandas(scanner, blob_mode, **kwargs)
+
     @abstractmethod
     def to_query_object(self) -> Query:
         """Return a serializable representation of the query
@@ -1357,6 +1624,9 @@ class LanceVectorQueryBuilder(LanceQueryBuilder):
             refine_factor=self._refine_factor,
             vector_column=self._vector_column,
             with_row_id=self._with_row_id,
+            with_row_address=self._with_row_address,
+            fragments=self._fragments,
+            fragment_ids=self._fragment_ids,
             offset=self._offset,
             fast_search=self._fast_search,
             ef=self._ef,
@@ -1559,6 +1829,9 @@ class LanceFtsQueryBuilder(LanceQueryBuilder):
             limit=self._limit,
             postfilter=self._postfilter,
             with_row_id=self._with_row_id,
+            with_row_address=self._with_row_address,
+            fragments=self._fragments,
+            fragment_ids=self._fragment_ids,
             full_text_query=FullTextSearchQuery(
                 query=self._query, columns=self._fts_columns
             ),
@@ -1629,6 +1902,9 @@ class LanceEmptyQueryBuilder(LanceQueryBuilder):
             filter=self._where,
             limit=self._limit,
             with_row_id=self._with_row_id,
+            with_row_address=self._with_row_address,
+            fragments=self._fragments,
+            fragment_ids=self._fragment_ids,
             offset=self._offset,
             order_by=self._order_by,
         )
@@ -2207,7 +2483,11 @@ class AsyncQueryBase(object):
     Base class for all async queries (take, scan, vector, fts, hybrid)
     """
 
-    def __init__(self, inner: Union[LanceQuery, LanceVectorQuery, LanceTakeQuery]):
+    def __init__(
+        self,
+        inner: Union[LanceQuery, LanceVectorQuery, LanceTakeQuery],
+        table: Optional["AsyncTable"] = None,
+    ):
         """
         Construct an AsyncQueryBase
 
@@ -2215,6 +2495,10 @@ class AsyncQueryBase(object):
         [AsyncTable.query][lancedb.table.AsyncTable.query] method to create a query.
         """
         self._inner = inner
+        self._table = table
+        self._with_row_address = None
+        self._fragments = None
+        self._fragment_ids = None
 
     def to_query_object(self) -> Query:
         """
@@ -2223,7 +2507,11 @@ class AsyncQueryBase(object):
         This is currently experimental but can be useful as the query object is pure
         python and more easily serializable.
         """
-        return Query.from_inner(self._inner.to_query_request())
+        query = Query.from_inner(self._inner.to_query_request())
+        query.with_row_address = self._with_row_address
+        query.fragments = self._fragments
+        query.fragment_ids = self._fragment_ids
+        return query
 
     def select(self, columns: Union[List[str], dict[str, str]]) -> Self:
         """
@@ -2280,6 +2568,27 @@ class AsyncQueryBase(object):
         self._inner.with_row_id()
         return self
 
+    def with_row_address(self, with_row_address: bool = True) -> Self:
+        """
+        Include the _rowaddr column in scanner-backed plain query results.
+        """
+        self._with_row_address = with_row_address
+        return self
+
+    def with_fragments(self, fragments: Any) -> Self:
+        """
+        Restrict scanner-backed plain query results to the given Lance fragments.
+        """
+        self._fragments = fragments
+        return self
+
+    def fragment_ids(self, fragment_ids: List[int]) -> Self:
+        """
+        Restrict scanner-backed plain query results to the given Lance fragment ids.
+        """
+        self._fragment_ids = fragment_ids
+        return self
+
     async def to_batches(
         self,
         *,
@@ -2357,6 +2666,8 @@ class AsyncQueryBase(object):
         self,
         flatten: Optional[Union[int, bool]] = None,
         timeout: Optional[timedelta] = None,
+        *,
+        blob_mode: BlobMode = "lazy",
         **kwargs,
     ) -> "pd.DataFrame":
         """
@@ -2390,13 +2701,63 @@ class AsyncQueryBase(object):
             The maximum time to wait for the query to complete.
             If not specified, no timeout is applied. If the query does not
             complete within the specified time, an error will be raised.
+        blob_mode: str, default "lazy"
+            Controls how blob columns are returned for plain scan queries.
+            Vector, FTS, hybrid, and other non-native query shapes keep the
+            existing Arrow conversion path and only support blob descriptions.
         **kwargs
             Forwarded to pyarrow.Table.to_pandas after query execution and
             optional flattening.
         """
-        return (
-            flatten_columns(await self.to_arrow(timeout=timeout), flatten)
-        ).to_pandas(**kwargs)
+        _validate_blob_mode(blob_mode)
+        if hasattr(self._inner, "output_schema"):
+            schema = await self.output_schema()
+            if _blob_mode_requires_native_pandas(blob_mode, schema):
+                native_error = None
+                if (flatten is None or blob_mode == "descriptions") and timeout is None:
+                    try:
+                        df = await self._plain_scan_to_pandas(
+                            blob_mode, flatten=flatten, **kwargs
+                        )
+                        if df is not None:
+                            return df
+                    except Exception as err:
+                        native_error = err
+                reason = (
+                    "this query shape cannot use Lance native pandas conversion"
+                    if native_error is None
+                    else str(native_error)
+                )
+                raise _unsupported_blob_pandas_error(reason) from native_error
+
+        tbl = flatten_columns(await self.to_arrow(timeout=timeout), flatten)
+        if _blob_mode_requires_native_pandas(blob_mode, tbl.schema):
+            raise _unsupported_blob_pandas_error(
+                "this query shape cannot use Lance native pandas conversion"
+            )
+        return tbl.to_pandas(**kwargs)
+
+    async def _plain_scan_to_pandas(
+        self,
+        blob_mode: BlobMode,
+        flatten: Optional[Union[int, bool]] = None,
+        **kwargs,
+    ) -> Optional["pd.DataFrame"]:
+        if self._table is None:
+            return None
+
+        query = self.to_query_object()
+        if not _query_is_plain_scan(query):
+            return None
+
+        dataset = await self._table._to_lance()
+        scanner = dataset.scanner(
+            **_scanner_kwargs_for_query(query, blob_mode, dataset)
+        )
+        if flatten is not None:
+            tbl = flatten_columns(_scanner_to_table(scanner), flatten)
+            return tbl.to_pandas(**kwargs)
+        return _scanner_to_pandas(scanner, blob_mode, **kwargs)
 
     async def to_polars(
         self,
@@ -2503,14 +2864,18 @@ class AsyncStandardQuery(AsyncQueryBase):
     Base class for "standard" async queries (all but take currently)
     """
 
-    def __init__(self, inner: Union[LanceQuery, LanceVectorQuery]):
+    def __init__(
+        self,
+        inner: Union[LanceQuery, LanceVectorQuery],
+        table: Optional["AsyncTable"] = None,
+    ):
         """
         Construct an AsyncStandardQuery
 
         This method is not intended to be called directly.  Instead, use the
         [AsyncTable.query][lancedb.table.AsyncTable.query] method to create a query.
         """
-        super().__init__(inner)
+        super().__init__(inner, table)
 
     def where(self, predicate: Union[str, Expr]) -> Self:
         """
@@ -2616,14 +2981,14 @@ class AsyncStandardQuery(AsyncQueryBase):
 
 
 class AsyncQuery(AsyncStandardQuery):
-    def __init__(self, inner: LanceQuery):
+    def __init__(self, inner: LanceQuery, table: Optional["AsyncTable"] = None):
         """
         Construct an AsyncQuery
 
         This method is not intended to be called directly.  Instead, use the
         [AsyncTable.query][lancedb.table.AsyncTable.query] method to create a query.
         """
-        super().__init__(inner)
+        super().__init__(inner, table)
         self._inner = inner
 
     @classmethod
@@ -2707,10 +3072,11 @@ class AsyncQuery(AsyncStandardQuery):
             new_self = self._inner.nearest_to(query_vectors[0])
             for v in query_vectors[1:]:
                 new_self.add_query_vector(v)
-            return AsyncVectorQuery(new_self)
+            return AsyncVectorQuery(new_self, self._table)
         else:
             return AsyncVectorQuery(
-                self._inner.nearest_to(AsyncQuery._query_vec_to_array(query_vector))
+                self._inner.nearest_to(AsyncQuery._query_vec_to_array(query_vector)),
+                self._table,
             )
 
     def nearest_to_text(
@@ -2743,17 +3109,18 @@ class AsyncQuery(AsyncStandardQuery):
 
         if isinstance(query, str):
             return AsyncFTSQuery(
-                self._inner.nearest_to_text({"query": query, "columns": columns})
+                self._inner.nearest_to_text({"query": query, "columns": columns}),
+                self._table,
             )
         # FullTextQuery object
-        return AsyncFTSQuery(self._inner.nearest_to_text({"query": query}))
+        return AsyncFTSQuery(self._inner.nearest_to_text({"query": query}), self._table)
 
 
 class AsyncFTSQuery(AsyncStandardQuery):
     """A query for full text search for LanceDB."""
 
-    def __init__(self, inner: LanceFTSQuery):
-        super().__init__(inner)
+    def __init__(self, inner: LanceFTSQuery, table: Optional["AsyncTable"] = None):
+        super().__init__(inner, table)
         self._inner = inner
         self._reranker = None
 
@@ -2835,10 +3202,11 @@ class AsyncFTSQuery(AsyncStandardQuery):
             new_self = self._inner.nearest_to(query_vectors[0])
             for v in query_vectors[1:]:
                 new_self.add_query_vector(v)
-            return AsyncHybridQuery(new_self)
+            return AsyncHybridQuery(new_self, self._table)
         else:
             return AsyncHybridQuery(
-                self._inner.nearest_to(AsyncQuery._query_vec_to_array(query_vector))
+                self._inner.nearest_to(AsyncQuery._query_vec_to_array(query_vector)),
+                self._table,
             )
 
     async def to_batches(
@@ -3029,7 +3397,7 @@ class AsyncVectorQueryBase:
 
 
 class AsyncVectorQuery(AsyncStandardQuery, AsyncVectorQueryBase):
-    def __init__(self, inner: LanceVectorQuery):
+    def __init__(self, inner: LanceVectorQuery, table: Optional["AsyncTable"] = None):
         """
         Construct an AsyncVectorQuery
 
@@ -3039,7 +3407,7 @@ class AsyncVectorQuery(AsyncStandardQuery, AsyncVectorQueryBase):
         a vector query.  Or you can use
         [AsyncTable.vector_search][lancedb.table.AsyncTable.vector_search]
         """
-        super().__init__(inner)
+        super().__init__(inner, table)
         self._inner = inner
         self._reranker = None
         self._query_string = None
@@ -3093,10 +3461,13 @@ class AsyncVectorQuery(AsyncStandardQuery, AsyncVectorQueryBase):
 
         if isinstance(query, str):
             return AsyncHybridQuery(
-                self._inner.nearest_to_text({"query": query, "columns": columns})
+                self._inner.nearest_to_text({"query": query, "columns": columns}),
+                self._table,
             )
         # FullTextQuery object
-        return AsyncHybridQuery(self._inner.nearest_to_text({"query": query}))
+        return AsyncHybridQuery(
+            self._inner.nearest_to_text({"query": query}), self._table
+        )
 
     async def to_batches(
         self,
@@ -3123,8 +3494,8 @@ class AsyncHybridQuery(AsyncStandardQuery, AsyncVectorQueryBase):
     in the `rerank` method to convert the scores to ranks and then normalize them.
     """
 
-    def __init__(self, inner: LanceHybridQuery):
-        super().__init__(inner)
+    def __init__(self, inner: LanceHybridQuery, table: Optional["AsyncTable"] = None):
+        super().__init__(inner, table)
         self._inner = inner
         self._norm = "score"
         self._reranker = RRFReranker()
@@ -3165,8 +3536,8 @@ class AsyncHybridQuery(AsyncStandardQuery, AsyncVectorQueryBase):
         max_batch_length: Optional[int] = None,
         timeout: Optional[timedelta] = None,
     ) -> AsyncRecordBatchReader:
-        fts_query = AsyncFTSQuery(self._inner.to_fts_query())
-        vec_query = AsyncVectorQuery(self._inner.to_vector_query())
+        fts_query = AsyncFTSQuery(self._inner.to_fts_query(), self._table)
+        vec_query = AsyncVectorQuery(self._inner.to_vector_query(), self._table)
 
         # save the row ID choice that was made on the query builder and force it
         # to actually fetch the row ids because we need this for reranking
@@ -3266,8 +3637,16 @@ class AsyncTakeQuery(AsyncQueryBase):
     Builder for parameterizing and executing take queries.
     """
 
-    def __init__(self, inner: LanceTakeQuery):
-        super().__init__(inner)
+    def __init__(self, inner: LanceTakeQuery, table: Optional["AsyncTable"] = None):
+        super().__init__(inner, table)
+
+    async def _plain_scan_to_pandas(
+        self,
+        blob_mode: BlobMode,
+        flatten: Optional[Union[int, bool]] = None,
+        **kwargs,
+    ) -> Optional["pd.DataFrame"]:
+        return None
 
 
 class BaseQueryBuilder(object):
@@ -3319,6 +3698,27 @@ class BaseQueryBuilder(object):
         self._inner.with_row_id()
         return self
 
+    def with_row_address(self, with_row_address: bool = True) -> Self:
+        """
+        Include the _rowaddr column in scanner-backed plain query results.
+        """
+        self._inner.with_row_address(with_row_address)
+        return self
+
+    def with_fragments(self, fragments: Any) -> Self:
+        """
+        Restrict scanner-backed plain query results to the given Lance fragments.
+        """
+        self._inner.with_fragments(fragments)
+        return self
+
+    def fragment_ids(self, fragment_ids: List[int]) -> Self:
+        """
+        Restrict scanner-backed plain query results to the given Lance fragment ids.
+        """
+        self._inner.fragment_ids(fragment_ids)
+        return self
+
     def output_schema(self) -> pa.Schema:
         """
         Return the output schema for the query
@@ -3400,6 +3800,8 @@ class BaseQueryBuilder(object):
         self,
         flatten: Optional[Union[int, bool]] = None,
         timeout: Optional[timedelta] = None,
+        *,
+        blob_mode: BlobMode = "lazy",
         **kwargs,
     ) -> "pd.DataFrame":
         """
@@ -3433,11 +3835,15 @@ class BaseQueryBuilder(object):
             The maximum time to wait for the query to complete.
             If not specified, no timeout is applied. If the query does not
             complete within the specified time, an error will be raised.
+        blob_mode: str, default "lazy"
+            Controls how blob columns are returned for plain scan queries.
         **kwargs
             Forwarded to pyarrow.Table.to_pandas after query execution and
             optional flattening.
         """
-        return LOOP.run(self._inner.to_pandas(flatten, timeout, **kwargs))
+        return LOOP.run(
+            self._inner.to_pandas(flatten, timeout, blob_mode=blob_mode, **kwargs)
+        )
 
     def to_polars(
         self,

python/python/lancedb/remote/db.py

@@ -3,6 +3,7 @@
 
 
 from datetime import timedelta
+import json
 import logging
 from concurrent.futures import ThreadPoolExecutor
 import sys
@@ -17,7 +18,7 @@ else:
 
 # Remove this import to fix circular dependency
 # from lancedb import connect_async
-from lancedb.remote import ClientConfig
+from lancedb.remote import ClientConfig, RetryConfig, TimeoutConfig, TlsConfig
 import pyarrow as pa
 
 from ..common import DATA
@@ -36,6 +37,64 @@ from ..table import Table
 from ..util import validate_table_name
 
 
+def _duration_seconds(value: Optional[timedelta]) -> Optional[float]:
+    return value.total_seconds() if value is not None else None
+
+
+def _timeout_config_to_dict(
+    config: Optional[TimeoutConfig],
+) -> Optional[dict[str, Any]]:
+    if config is None:
+        return None
+    return {
+        "timeout": _duration_seconds(config.timeout),
+        "connect_timeout": _duration_seconds(config.connect_timeout),
+        "read_timeout": _duration_seconds(config.read_timeout),
+        "pool_idle_timeout": _duration_seconds(config.pool_idle_timeout),
+    }
+
+
+def _retry_config_to_dict(config: RetryConfig) -> dict[str, Any]:
+    return {
+        "retries": config.retries,
+        "connect_retries": config.connect_retries,
+        "read_retries": config.read_retries,
+        "backoff_factor": config.backoff_factor,
+        "backoff_jitter": config.backoff_jitter,
+        "statuses": config.statuses,
+    }
+
+
+def _tls_config_to_dict(config: Optional[TlsConfig]) -> Optional[dict[str, Any]]:
+    if config is None:
+        return None
+    return {
+        "cert_file": config.cert_file,
+        "key_file": config.key_file,
+        "ssl_ca_cert": config.ssl_ca_cert,
+        "assert_hostname": config.assert_hostname,
+    }
+
+
+def _client_config_to_dict(config: ClientConfig) -> dict[str, Any]:
+    if config.header_provider is not None:
+        raise ValueError(
+            "Cannot serialize a remote connection with a header_provider. "
+            "Use static api_key/extra_headers or provide a worker-side "
+            "connection factory instead."
+        )
+    return {
+        "user_agent": config.user_agent,
+        "retry_config": _retry_config_to_dict(config.retry_config),
+        "timeout_config": _timeout_config_to_dict(config.timeout_config),
+        "extra_headers": config.extra_headers,
+        "id_delimiter": config.id_delimiter,
+        "tls_config": _tls_config_to_dict(config.tls_config),
+        "header_provider": None,
+        "user_id": config.user_id,
+    }
+
+
 class RemoteDBConnection(DBConnection):
     """A connection to a remote LanceDB database."""
 
@@ -50,6 +109,7 @@ class RemoteDBConnection(DBConnection):
         connection_timeout: Optional[float] = None,
         read_timeout: Optional[float] = None,
         storage_options: Optional[Dict[str, str]] = None,
+        read_consistency_interval: Optional[timedelta] = None,
     ):
         """Connect to a remote LanceDB database."""
         if isinstance(client_config, dict):
@@ -88,6 +148,11 @@ class RemoteDBConnection(DBConnection):
         parsed = urlparse(db_url)
         if parsed.scheme != "db":
             raise ValueError(f"Invalid scheme: {parsed.scheme}, only accepts db://")
+        self.db_url = db_url
+        self.api_key = api_key
+        self.region = region
+        self.host_override = host_override
+        self.storage_options = storage_options
         self.db_name = parsed.netloc
 
         self.client_config = client_config
@@ -103,12 +168,27 @@ class RemoteDBConnection(DBConnection):
                 host_override=host_override,
                 client_config=client_config,
                 storage_options=storage_options,
+                read_consistency_interval=read_consistency_interval,
             )
         )
 
     def __repr__(self) -> str:
         return f"RemoteConnect(name={self.db_name})"
 
+    @override
+    def serialize(self) -> str:
+        return json.dumps(
+            {
+                "connection_type": "remote",
+                "db_url": self.db_url,
+                "api_key": self.api_key,
+                "region": self.region,
+                "host_override": self.host_override,
+                "client_config": _client_config_to_dict(self.client_config),
+                "storage_options": self.storage_options,
+            }
+        )
+
     @override
     def list_namespaces(
         self,
@@ -303,6 +383,8 @@ class RemoteDBConnection(DBConnection):
         namespace_path: Optional[List[str]] = None,
         storage_options: Optional[Dict[str, str]] = None,
         index_cache_size: Optional[int] = None,
+        branch: Optional[str] = None,
+        version: Optional[int] = None,
     ) -> Table:
         """Open a Lance Table in the database.
 
@@ -313,6 +395,14 @@ class RemoteDBConnection(DBConnection):
         namespace_path: List[str], optional
             The namespace to open the table from.
             None or empty list represents root namespace.
+        branch: str, optional
+            Branching is not yet supported on remote tables, so only the
+            default branch is accepted (``None`` or ``"main"``); any other
+            value raises ``NotImplementedError``.
+        version: int, optional
+            If provided, open the table pinned to this version, producing a
+            read-only handle. Call ``checkout_latest`` to return to a writable
+            state.
 
         Returns
         -------
@@ -320,6 +410,11 @@ class RemoteDBConnection(DBConnection):
         """
         from .table import RemoteTable
 
+        # Remote supports version time-travel but not branches: reject a non-main
+        # branch, but allow a version-only open (or "main").
+        if branch is not None and branch != "main":
+            raise NotImplementedError("branching is not yet supported on remote tables")
+
         if namespace_path is None:
             namespace_path = []
         if index_cache_size is not None:
@@ -329,7 +424,15 @@ class RemoteDBConnection(DBConnection):
             )
 
         table = LOOP.run(self._conn.open_table(name, namespace_path=namespace_path))
-        return RemoteTable(table, self.db_name)
+        tbl = RemoteTable(
+            table,
+            self.db_name,
+            connection_state=self.serialize,
+            namespace_path=namespace_path,
+        )
+        if version is not None:
+            tbl.checkout(version)
+        return tbl
 
     def clone_table(
         self,
@@ -378,7 +481,12 @@ class RemoteDBConnection(DBConnection):
                 is_shallow=is_shallow,
             )
         )
-        return RemoteTable(table, self.db_name)
+        return RemoteTable(
+            table,
+            self.db_name,
+            connection_state=self.serialize,
+            namespace_path=target_namespace_path,
+        )
 
     @override
     def create_table(
@@ -523,7 +631,12 @@ class RemoteDBConnection(DBConnection):
                 fill_value=fill_value,
             )
         )
-        return RemoteTable(table, self.db_name)
+        return RemoteTable(
+            table,
+            self.db_name,
+            connection_state=self.serialize,
+            namespace_path=namespace_path,
+        )
 
     @override
     def drop_table(self, name: str, namespace_path: Optional[List[str]] = None):

python/python/lancedb/remote/errors.py

@@ -27,6 +27,9 @@ class LanceDBClientError(RuntimeError):
         self.request_id = request_id
         self.status_code = status_code
 
+    def __reduce__(self) -> tuple[type, tuple]:
+        return (self.__class__, (str(self), self.request_id, self.status_code))
+
 
 class HttpError(LanceDBClientError):
     """An error that occurred during an HTTP request.
@@ -101,3 +104,19 @@ class RetryError(LanceDBClientError):
         self.max_request_failures = max_request_failures
         self.max_connect_failures = max_connect_failures
         self.max_read_failures = max_read_failures
+
+    def __reduce__(self) -> tuple[type, tuple]:
+        return (
+            self.__class__,
+            (
+                str(self),
+                self.request_id,
+                self.request_failures,
+                self.connect_failures,
+                self.read_failures,
+                self.max_request_failures,
+                self.max_connect_failures,
+                self.max_read_failures,
+                self.status_code,
+            ),
+        )

python/python/lancedb/remote/table.py

@@ -2,15 +2,30 @@
 # SPDX-FileCopyrightText: Copyright The LanceDB Authors
 
 from datetime import timedelta
+import deprecation
 import logging
 from functools import cached_property
-from typing import Any, Callable, Dict, Iterable, List, Optional, Union, Literal
+import os
+from typing import (
+    Any,
+    Callable,
+    Dict,
+    Iterable,
+    List,
+    Optional,
+    Union,
+    Literal,
+    overload,
+)
 import warnings
 
+from lancedb import __version__
+
 from lancedb._lancedb import (
     AddColumnsResult,
     AddResult,
     AlterColumnsResult,
+    UpdateFieldMetadataResult,
     DeleteResult,
     DropColumnsResult,
     IndexConfig,
@@ -32,6 +47,7 @@ from lancedb.index import (
     LabelList,
 )
 from lancedb.remote.db import LOOP
+from lancedb.table import IndexConfigType, KNOWN_METRICS
 import pyarrow as pa
 
 from lancedb.common import DATA, VEC, VECTOR_COLUMN_NAME
@@ -49,14 +65,80 @@ class RemoteTable(Table):
         self,
         table: AsyncTable,
         db_name: str,
+        *,
+        connection_state: Optional[Union[str, Callable[[], str]]] = None,
+        namespace_path: Optional[List[str]] = None,
     ):
-        self._table = table
+        self._table_handle = table
+        self._name = table.name
         self.db_name = db_name
+        self._connection_state = connection_state
+        self._namespace_path = list(namespace_path or [])
+        self._checkout_version: Optional[int] = None
+        self._pid = os.getpid()
+
+    def _serialized_connection_state(self) -> str:
+        if self._connection_state is None:
+            raise RuntimeError(
+                "Cannot reopen this remote table because it does not carry "
+                "serialized connection state"
+            )
+        if callable(self._connection_state):
+            self._connection_state = self._connection_state()
+        return self._connection_state
+
+    @property
+    def _table(self) -> AsyncTable:
+        self._ensure_open()
+        assert self._table_handle is not None
+        return self._table_handle
+
+    @_table.setter
+    def _table(self, table: AsyncTable) -> None:
+        self._table_handle = table
+        self._name = table.name
+        self._pid = os.getpid()
+
+    def _ensure_open(self) -> None:
+        pid = os.getpid()
+        if self._table_handle is not None and self._pid == pid:
+            return
+
+        # Pickle clears the handle; fork inherits a handle created in the
+        # parent process. In both cases reopen before touching the Rust client.
+        from lancedb import deserialize_conn
+
+        db = deserialize_conn(self._serialized_connection_state(), for_worker=True)
+        table = db.open_table(self._name, namespace_path=self._namespace_path)
+        if self._checkout_version is not None:
+            table.checkout(self._checkout_version)
+
+        self._table_handle = table._table
+        self.db_name = table.db_name
+        self._pid = pid
+
+    def __getstate__(self) -> dict:
+        return {
+            "connection_state": self._serialized_connection_state(),
+            "db_name": self.db_name,
+            "name": self.name,
+            "namespace_path": self._namespace_path,
+            "checkout_version": self._checkout_version,
+        }
+
+    def __setstate__(self, state: dict) -> None:
+        self._table_handle = None
+        self._name = state["name"]
+        self.db_name = state["db_name"]
+        self._connection_state = state["connection_state"]
+        self._namespace_path = state["namespace_path"]
+        self._checkout_version = state["checkout_version"]
+        self._pid = None
 
     @property
     def name(self) -> str:
         """The name of the table"""
-        return self._table.name
+        return self._name
 
     def __repr__(self) -> str:
         return f"RemoteTable({self.db_name}.{self.name})"
@@ -106,13 +188,19 @@ class RemoteTable(Table):
         raise NotImplementedError("to_pandas() is not yet supported on LanceDB cloud.")
 
     def checkout(self, version: Union[int, str]):
-        return LOOP.run(self._table.checkout(version))
+        result = LOOP.run(self._table.checkout(version))
+        self._checkout_version = self.version
+        return result
 
     def checkout_latest(self):
-        return LOOP.run(self._table.checkout_latest())
+        result = LOOP.run(self._table.checkout_latest())
+        self._checkout_version = None
+        return result
 
     def restore(self, version: Optional[Union[int, str]] = None):
-        return LOOP.run(self._table.restore(version))
+        result = LOOP.run(self._table.restore(version))
+        self._checkout_version = None
+        return result
 
     def list_indices(self) -> Iterable[IndexConfig]:
         """List all the indices on the table"""
@@ -122,6 +210,11 @@ class RemoteTable(Table):
         """List all the stats of a specified index"""
         return LOOP.run(self._table.index_stats(index_uuid))
 
+    @deprecation.deprecated(
+        deprecated_in="0.25.0",
+        current_version=__version__,
+        details="Use create_index() with config=BTree()/Bitmap()/LabelList() instead.",
+    )
     def create_scalar_index(
         self,
         column: str,
@@ -131,7 +224,12 @@ class RemoteTable(Table):
         wait_timeout: Optional[timedelta] = None,
         name: Optional[str] = None,
     ):
-        """Creates a scalar index
+        """Creates a scalar index.
+
+        .. deprecated:: 0.25.0
+            Use :meth:`create_index` with a BTree, Bitmap, or LabelList config instead.
+            Example: ``table.create_index("column", config=BTree())``
+
         Parameters
         ----------
         column : str
@@ -162,6 +260,11 @@ class RemoteTable(Table):
             )
         )
 
+    @deprecation.deprecated(
+        deprecated_in="0.25.0",
+        current_version=__version__,
+        details="Use create_index() with config=FTS() instead.",
+    )
     def create_fts_index(
         self,
         column: str,
@@ -182,6 +285,12 @@ class RemoteTable(Table):
         prefix_only: bool = False,
         name: Optional[str] = None,
     ):
+        """Create a full-text search index on a column.
+
+        .. deprecated:: 0.25.0
+            Use :meth:`create_index` with an FTS config instead.
+            Example: ``table.create_index("text_column", config=FTS())``
+        """
         config = FTS(
             with_position=with_position,
             base_tokenizer=base_tokenizer,
@@ -205,9 +314,43 @@ class RemoteTable(Table):
             )
         )
 
+    # New unified API overload
+    @overload
     def create_index(
         self,
-        metric="l2",
+        column: str,
+        /,
+        *,
+        config: IndexConfigType,
+        wait_timeout: Optional[timedelta] = ...,
+        name: Optional[str] = ...,
+        train: bool = ...,
+    ) -> None: ...
+
+    # Legacy API overload (deprecated)
+    @overload
+    def create_index(
+        self,
+        metric: Literal["l2", "cosine", "dot", "hamming"] = ...,
+        vector_column_name: str = ...,
+        index_cache_size: Optional[int] = ...,
+        num_partitions: Optional[int] = ...,
+        num_sub_vectors: Optional[int] = ...,
+        replace: Optional[bool] = ...,
+        accelerator: Optional[str] = ...,
+        index_type: Literal[
+            "VECTOR", "IVF_FLAT", "IVF_SQ", "IVF_PQ", "IVF_HNSW_SQ", "IVF_HNSW_PQ"
+        ] = ...,
+        wait_timeout: Optional[timedelta] = ...,
+        *,
+        num_bits: int = ...,
+        name: Optional[str] = ...,
+        train: bool = ...,
+    ) -> None: ...
+
+    def create_index(
+        self,
+        metric: str = "l2",
         vector_column_name: str = VECTOR_COLUMN_NAME,
         index_cache_size: Optional[int] = None,
         num_partitions: Optional[int] = None,
@@ -218,89 +361,113 @@ class RemoteTable(Table):
         wait_timeout: Optional[timedelta] = None,
         *,
         num_bits: int = 8,
+        config: Optional[IndexConfigType] = None,
         name: Optional[str] = None,
         train: bool = True,
     ):
-        """Create an index on the table.
+        """Create an index on a column.
 
-        Parameters
-        ----------
-        metric : str
-            The metric to use for the index. Default is "l2".
-        vector_column_name : str
-            The name of the vector column. Default is "vector".
+        This method supports both the new unified API and the legacy API
+        for backwards compatibility. The new API takes the column name as the
+        first positional argument and an index configuration object via
+        ``config``; the legacy API takes the distance metric as the first
+        argument plus separate ``vector_column_name`` / ``num_partitions`` /
+        etc. parameters, and emits a ``DeprecationWarning``.
 
         Examples
         --------
-        >>> import lancedb
-        >>> import uuid
-        >>> from lancedb.schema import vector
-        >>> db = lancedb.connect("db://...", api_key="...", # doctest: +SKIP
-        ...                      region="...") # doctest: +SKIP
-        >>> table_name = uuid.uuid4().hex
-        >>> schema = pa.schema(
-        ...     [
-        ...             pa.field("id", pa.uint32(), False),
-        ...            pa.field("vector", vector(128), False),
-        ...             pa.field("s", pa.string(), False),
-        ...     ]
+        New API (recommended):
+
+        >>> table.create_index(  # doctest: +SKIP
+        ...     "vector", config=IvfPq(distance_type="l2")
         ... )
-        >>> table = db.create_table( # doctest: +SKIP
-        ...     table_name, # doctest: +SKIP
-        ...     schema=schema, # doctest: +SKIP
+        >>> table.create_index("category", config=BTree())  # doctest: +SKIP
+        >>> table.create_index("content", config=FTS())  # doctest: +SKIP
+
+        Legacy API (deprecated):
+
+        >>> table.create_index(  # doctest: +SKIP
+        ...     "l2", vector_column_name="vector"
         ... )
-        >>> table.create_index("l2", "vector") # doctest: +SKIP
         """
+        # Detect whether this is a legacy API call
+        is_legacy = self._is_legacy_create_index_call(
+            metric,
+            config,
+            num_partitions,
+            num_sub_vectors,
+            vector_column_name,
+            accelerator,
+            index_cache_size,
+            replace,
+        )
 
-        if accelerator is not None:
-            logging.warning(
-                "GPU accelerator is not yet supported on LanceDB cloud."
-                "If you have 100M+ vectors to index,"
-                "please contact us at contact@lancedb.com"
-            )
-        if replace is not None:
-            logging.warning(
-                "replace is not supported on LanceDB cloud."
-                "Existing indexes will always be replaced."
+        if is_legacy:
+            warnings.warn(
+                "The create_index() API with metric/num_partitions parameters is "
+                "deprecated and will be removed in a future version. "
+                "Please migrate to the new unified API:\n"
+                "  # Old (deprecated):\n"
+                "  table.create_index('l2', vector_column_name='my_vector')\n"
+                "  # New (recommended):\n"
+                "  table.create_index('my_vector', config=IvfPq(distance_type='l2'))",
+                DeprecationWarning,
+                stacklevel=2,
             )
 
-        index_type = index_type.upper()
-        if index_type == "VECTOR" or index_type == "IVF_PQ":
-            config = IvfPq(
-                distance_type=metric,
-                num_partitions=num_partitions,
-                num_sub_vectors=num_sub_vectors,
-                num_bits=num_bits,
-            )
-        elif index_type == "IVF_RQ":
-            config = IvfRq(
-                distance_type=metric,
-                num_partitions=num_partitions,
-                num_bits=num_bits,
-            )
-        elif index_type == "IVF_SQ":
-            config = IvfSq(distance_type=metric, num_partitions=num_partitions)
-        elif index_type == "IVF_HNSW_PQ":
-            raise ValueError(
-                "IVF_HNSW_PQ is not supported on LanceDB cloud."
-                "Please use IVF_HNSW_SQ instead."
-            )
-        elif index_type == "IVF_HNSW_SQ":
-            config = HnswSq(distance_type=metric, num_partitions=num_partitions)
-        elif index_type == "IVF_HNSW_FLAT":
-            config = HnswFlat(distance_type=metric, num_partitions=num_partitions)
-        elif index_type == "IVF_FLAT":
-            config = IvfFlat(distance_type=metric, num_partitions=num_partitions)
+            column = vector_column_name
+
+            if accelerator is not None:
+                logging.warning(
+                    "GPU accelerator is not yet supported on LanceDB cloud."
+                    "If you have 100M+ vectors to index,"
+                    "please contact us at contact@lancedb.com"
+                )
+            if replace is not None:
+                logging.warning(
+                    "replace is not supported on LanceDB cloud."
+                    "Existing indexes will always be replaced."
+                )
+
+            idx_type = index_type.upper()
+            if idx_type == "VECTOR" or idx_type == "IVF_PQ":
+                config = IvfPq(
+                    distance_type=metric,
+                    num_partitions=num_partitions,
+                    num_sub_vectors=num_sub_vectors,
+                    num_bits=num_bits,
+                )
+            elif idx_type == "IVF_RQ":
+                config = IvfRq(
+                    distance_type=metric,
+                    num_partitions=num_partitions,
+                    num_bits=num_bits,
+                )
+            elif idx_type == "IVF_SQ":
+                config = IvfSq(distance_type=metric, num_partitions=num_partitions)
+            elif idx_type == "IVF_HNSW_PQ":
+                raise ValueError(
+                    "IVF_HNSW_PQ is not supported on LanceDB cloud."
+                    "Please use IVF_HNSW_SQ instead."
+                )
+            elif idx_type == "IVF_HNSW_SQ":
+                config = HnswSq(distance_type=metric, num_partitions=num_partitions)
+            elif idx_type == "IVF_HNSW_FLAT":
+                config = HnswFlat(distance_type=metric, num_partitions=num_partitions)
+            elif idx_type == "IVF_FLAT":
+                config = IvfFlat(distance_type=metric, num_partitions=num_partitions)
+            else:
+                raise ValueError(
+                    f"Unknown vector index type: {idx_type}. Valid options are"
+                    " 'IVF_FLAT', 'IVF_PQ', 'IVF_RQ', 'IVF_SQ',"
+                    " 'IVF_HNSW_PQ', 'IVF_HNSW_SQ', 'IVF_HNSW_FLAT'"
+                )
         else:
-            raise ValueError(
-                f"Unknown vector index type: {index_type}. Valid options are"
-                " 'IVF_FLAT', 'IVF_PQ', 'IVF_RQ', 'IVF_SQ',"
-                " 'IVF_HNSW_PQ', 'IVF_HNSW_SQ', 'IVF_HNSW_FLAT'"
-            )
+            column = metric
 
         LOOP.run(
             self._table.create_index(
-                vector_column_name,
+                column,
                 config=config,
                 wait_timeout=wait_timeout,
                 name=name,
@@ -308,6 +475,37 @@ class RemoteTable(Table):
             )
         )
 
+    def _is_legacy_create_index_call(
+        self,
+        first_arg: str,
+        config: Optional[IndexConfigType],
+        num_partitions: Optional[int],
+        num_sub_vectors: Optional[int],
+        vector_column_name: str,
+        accelerator: Optional[str],
+        index_cache_size: Optional[int],
+        replace: Optional[bool],
+    ) -> bool:
+        """Detect if this is a legacy create_index call."""
+        if config is not None:
+            return False
+        if any(
+            x is not None
+            for x in (
+                num_partitions,
+                num_sub_vectors,
+                accelerator,
+                index_cache_size,
+                replace,
+            )
+        ):
+            return True
+        if vector_column_name != VECTOR_COLUMN_NAME:
+            return True
+        if first_arg.lower() in KNOWN_METRICS:
+            return True
+        return False
+
     def add(
         self,
         data: DATA,
@@ -653,6 +851,11 @@ class RemoteTable(Table):
     ) -> AlterColumnsResult:
         return LOOP.run(self._table.alter_columns(*alterations))
 
+    def update_field_metadata(
+        self, *updates: dict[str, Any]
+    ) -> UpdateFieldMetadataResult:
+        return LOOP.run(self._table.update_field_metadata(*updates))
+
     def drop_columns(self, columns: Iterable[str]) -> DropColumnsResult:
         return LOOP.run(self._table.drop_columns(columns))
 
@@ -668,6 +871,10 @@ class RemoteTable(Table):
         """Not supported on LanceDB Cloud."""
         return LOOP.run(self._table.unset_lsm_write_spec())
 
+    def close_lsm_writers(self) -> None:
+        """No-op on LanceDB Cloud (no local shard writers)."""
+        return LOOP.run(self._table.close_lsm_writers())
+
     def drop_index(self, index_name: str):
         return LOOP.run(self._table.drop_index(index_name))
 

python/python/lancedb/rerankers/linear_combination.py

@@ -102,8 +102,15 @@ class LinearCombinationReranker(Reranker):
 
         combined_list = []
         for row_id, result in results.items():
+            # Convert vector distance to a relevance score in [0, 1] where
+            # higher is better.  Missing vector entries are penalised with
+            # `_invert_score(fill)` = 1 - fill (= 0.0 for the default fill=1).
             vector_score = self._invert_score(result.get("_distance", fill))
-            fts_score = result.get("_score", fill)
+            # FTS scores (BM25) are already in a "higher = more relevant" space.
+            # Missing FTS entries are penalised symmetrically: we use
+            # `1 - fill` so that the same `fill` value drives both missing-vector
+            # and missing-FTS penalties in the same direction.
+            fts_score = result.get("_score", 1 - fill)
             result["_relevance_score"] = self._combine_score(vector_score, fts_score)
             combined_list.append(result)
 
@@ -123,8 +130,12 @@ class LinearCombinationReranker(Reranker):
         return tbl
 
     def _combine_score(self, vector_score, fts_score):
-        # these scores represent distance
-        return 1 - (self.weight * vector_score + (1 - self.weight) * fts_score)
+        # Both vector_score (inverted distance) and fts_score are in a
+        # "higher = more relevant" space.  A straight weighted average gives
+        # higher _relevance_score to better matches, as expected.
+        # Previously this returned `1 - (...)` which inverted the final
+        # ranking so that the *least* relevant document ranked first.
+        return self.weight * vector_score + (1 - self.weight) * fts_score
 
     def _invert_score(self, dist: float):
         # Invert the score between relevance and distance

python/python/lancedb/rerankers/mrr.py

@@ -125,6 +125,9 @@ class MRRReranker(Reranker):
         This cannot reuse rerank_hybrid because MRR semantics require treating
         each vector result as a separate ranking system.
         """
+        if not vector_results:
+            raise ValueError("vector_results must not be empty")
+
         if not all(isinstance(v, type(vector_results[0])) for v in vector_results):
             raise ValueError(
                 "All elements in vector_results should be of the same type"

python/python/lancedb/rerankers/rrf.py

@@ -82,6 +82,9 @@ class RRFReranker(Reranker):
         results from multiple vector searches as it doesn't support reranking
         vector results individually.
         """
+        if not vector_results:
+            raise ValueError("vector_results must not be empty")
+
         # Make sure all elements are of the same type
         if not all(isinstance(v, type(vector_results[0])) for v in vector_results):
             raise ValueError(

python/python/lancedb/util.py

@@ -385,6 +385,21 @@ def _(value: np.ndarray):
     return value_to_sql(value.tolist())
 
 
+@value_to_sql.register(np.bool_)
+def _(value: np.bool_):
+    return value_to_sql(bool(value))
+
+
+@value_to_sql.register(np.integer)
+def _(value: np.integer):
+    return value_to_sql(int(value))
+
+
+@value_to_sql.register(np.floating)
+def _(value: np.floating):
+    return value_to_sql(float(value))
+
+
 def deprecated(func):
     """This is a decorator which can be used to mark functions
     as deprecated. It will result in a warning being emitted

python/python/tests/docs/test_merge_insert.py

@@ -57,7 +57,7 @@ async def test_upsert_async(mem_db_async):
     await table.count_rows()  # 3
     res
     # MergeResult(version=2, num_updated_rows=1,
-    # num_inserted_rows=1, num_deleted_rows=0)
+    # num_inserted_rows=1, num_deleted_rows=0, num_rows=2)
     # --8<-- [end:upsert_basic_async]
     assert await table.count_rows() == 3
     assert res.version == 2
@@ -86,7 +86,7 @@ def test_insert_if_not_exists(mem_db):
     table.count_rows()  # 3
     res
     # MergeResult(version=2, num_updated_rows=0,
-    # num_inserted_rows=1, num_deleted_rows=0)
+    # num_inserted_rows=1, num_deleted_rows=0, num_rows=1)
     # --8<-- [end:insert_if_not_exists]
     assert table.count_rows() == 3
     assert res.version == 2
@@ -116,7 +116,7 @@ async def test_insert_if_not_exists_async(mem_db_async):
     await table.count_rows()  # 3
     res
     # MergeResult(version=2, num_updated_rows=0,
-    # num_inserted_rows=1, num_deleted_rows=0)
+    # num_inserted_rows=1, num_deleted_rows=0, num_rows=1)
     # --8<-- [end:insert_if_not_exists]
     assert await table.count_rows() == 3
     assert res.version == 2
@@ -150,7 +150,7 @@ def test_replace_range(mem_db):
     table.count_rows("doc_id = 1")  # 1
     res
     # MergeResult(version=2, num_updated_rows=1,
-    # num_inserted_rows=0, num_deleted_rows=1)
+    # num_inserted_rows=0, num_deleted_rows=1, num_rows=1)
     # --8<-- [end:insert_if_not_exists]
     assert table.count_rows("doc_id = 1") == 1
     assert res.version == 2
@@ -185,7 +185,7 @@ async def test_replace_range_async(mem_db_async):
     await table.count_rows("doc_id = 1")  # 1
     res
     # MergeResult(version=2, num_updated_rows=1,
-    # num_inserted_rows=0, num_deleted_rows=1)
+    # num_inserted_rows=0, num_deleted_rows=1, num_rows=1)
     # --8<-- [end:insert_if_not_exists]
     assert await table.count_rows("doc_id = 1") == 1
     assert res.version == 2

python/python/tests/test_db.py

@@ -466,7 +466,8 @@ async def test_create_table_v2_manifest_paths_async(tmp_path):
     assert await tbl.uses_v2_manifest_paths()
     manifests_dir = tmp_path / "test_v2_manifest_paths.lance" / "_versions"
     for manifest in os.listdir(manifests_dir):
-        assert re.match(r"\d{20}\.manifest", manifest)
+        if manifest.endswith(".manifest"):
+            assert re.match(r"\d{20}\.manifest", manifest)
 
     # Start a table in V1 mode then migrate
     tbl = await db_no_v2_paths.create_table(
@@ -476,13 +477,15 @@ async def test_create_table_v2_manifest_paths_async(tmp_path):
     assert not await tbl.uses_v2_manifest_paths()
     manifests_dir = tmp_path / "test_v2_migration.lance" / "_versions"
     for manifest in os.listdir(manifests_dir):
-        assert re.match(r"\d\.manifest", manifest)
+        if manifest.endswith(".manifest"):
+            assert re.match(r"\d\.manifest", manifest)
 
     await tbl.migrate_manifest_paths_v2()
     assert await tbl.uses_v2_manifest_paths()
 
     for manifest in os.listdir(manifests_dir):
-        assert re.match(r"\d{20}\.manifest", manifest)
+        if manifest.endswith(".manifest"):
+            assert re.match(r"\d{20}\.manifest", manifest)
 
 
 @pytest.mark.asyncio

python/python/tests/test_errors.py

@@ -0,0 +1,56 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright The LanceDB Authors
+
+import pickle
+
+from lancedb.remote.errors import HttpError, LanceDBClientError, RetryError
+
+
+def test_pickle_lancedb_client_error():
+    err = LanceDBClientError("something went wrong", "req-123", 400)
+    restored = pickle.loads(pickle.dumps(err))
+    assert str(restored) == "something went wrong"
+    assert restored.request_id == "req-123"
+    assert restored.status_code == 400
+
+
+def test_pickle_lancedb_client_error_no_status_code():
+    err = LanceDBClientError("fail", "req-456")
+    restored = pickle.loads(pickle.dumps(err))
+    assert str(restored) == "fail"
+    assert restored.request_id == "req-456"
+    assert restored.status_code is None
+
+
+def test_pickle_http_error():
+    err = HttpError("not found", "req-789", 404)
+    restored = pickle.loads(pickle.dumps(err))
+    assert isinstance(restored, HttpError)
+    assert str(restored) == "not found"
+    assert restored.request_id == "req-789"
+    assert restored.status_code == 404
+
+
+def test_pickle_retry_error():
+    err = RetryError(
+        "max retries exceeded",
+        "req-abc",
+        request_failures=3,
+        connect_failures=1,
+        read_failures=2,
+        max_request_failures=5,
+        max_connect_failures=3,
+        max_read_failures=3,
+        status_code=503,
+    )
+    restored = pickle.loads(pickle.dumps(err))
+    assert isinstance(restored, RetryError)
+    assert str(restored) == "max retries exceeded"
+    assert restored.request_id == "req-abc"
+    assert restored.request_failures == 3
+    assert restored.connect_failures == 1
+    assert restored.read_failures == 2
+    assert restored.max_request_failures == 5
+    assert restored.max_connect_failures == 3
+    assert restored.max_read_failures == 3
+    assert restored.status_code == 503

python/python/tests/test_fts.py

@@ -215,11 +215,12 @@ def test_reject_legacy_tantivy_index(table):
 
 @pytest.mark.parametrize("with_position", [True, False])
 def test_create_inverted_index(table, with_position):
-    table.create_fts_index(
-        "text",
-        with_position=with_position,
-        name="custom_fts_index",
-    )
+    with pytest.warns(DeprecationWarning, match="create_fts_index"):
+        table.create_fts_index(
+            "text",
+            with_position=with_position,
+            name="custom_fts_index",
+        )
     indices = table.list_indices()
     fts_indices = [i for i in indices if i.index_type == "FTS"]
     assert any(i.name == "custom_fts_index" for i in fts_indices)

python/python/tests/test_index.py

@@ -162,12 +162,13 @@ async def test_create_bitmap_index(some_table: AsyncTable):
     await some_table.create_index("data", config=Bitmap())
     indices = await some_table.list_indices()
     assert len(indices) == 3
+    # list_indices returns indices in alphabetical order by name
     assert indices[0].index_type == "Bitmap"
-    assert indices[0].columns == ["id"]
+    assert indices[0].columns == ["data"]
     assert indices[1].index_type == "Bitmap"
-    assert indices[1].columns == ["is_active"]
+    assert indices[1].columns == ["id"]
     assert indices[2].index_type == "Bitmap"
-    assert indices[2].columns == ["data"]
+    assert indices[2].columns == ["is_active"]
 
     index_name = indices[0].name
     stats = await some_table.index_stats(index_name)
@@ -225,7 +226,6 @@ async def test_create_vector_index(some_table: AsyncTable):
     assert stats.num_indexed_rows == await some_table.count_rows()
     assert stats.num_unindexed_rows == 0
     assert stats.num_indices == 1
-    assert stats.loss >= 0.0
 
 
 @pytest.mark.asyncio
@@ -249,7 +249,6 @@ async def test_create_4bit_ivfpq_index(some_table: AsyncTable):
     assert stats.num_indexed_rows == await some_table.count_rows()
     assert stats.num_unindexed_rows == 0
     assert stats.num_indices == 1
-    assert stats.loss >= 0.0
 
 
 @pytest.mark.asyncio

python/python/tests/test_lsm_write_spec.py

@@ -40,16 +40,6 @@ def _make_table(tmp_path):
 def test_set_lsm_write_spec_validates(tmp_path):
     _db, table = _make_table(tmp_path)
 
-    # No PK set yet.
-    with pytest.raises(Exception, match="primary key"):
-        table.set_lsm_write_spec(LsmWriteSpec.bucket("id", 4))
-
-    table.set_unenforced_primary_key("id")
-
-    # Column mismatch.
-    with pytest.raises(Exception, match="match"):
-        table.set_lsm_write_spec(LsmWriteSpec.bucket("v", 4))
-
     # Out-of-range num_buckets.
     with pytest.raises(Exception, match="num_buckets"):
         table.set_lsm_write_spec(LsmWriteSpec.bucket("id", 0))
@@ -70,7 +60,6 @@ def test_unset_lsm_write_spec(tmp_path):
         table.unset_lsm_write_spec()
 
     # Install a spec, then remove it; afterwards a fresh spec can be set.
-    table.set_unenforced_primary_key("id")
     table.set_lsm_write_spec(LsmWriteSpec.bucket("id", 4))
     table.unset_lsm_write_spec()
     # A second unset errors — there is no spec left to remove.

python/python/tests/test_merge_insert_lsm.py

@@ -0,0 +1,196 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright The LanceDB Authors
+
+"""Tests for the MemWAL LSM ``merge_insert`` dispatch."""
+
+from datetime import timedelta
+
+import lancedb
+import pyarrow as pa
+import pytest
+from lancedb._lancedb import LsmWriteSpec
+
+SCHEMA = pa.schema(
+    [
+        pa.field("id", pa.int64(), nullable=False),
+        pa.field("value", pa.int64(), nullable=False),
+    ]
+)
+
+REGION_SCHEMA = pa.schema(
+    [
+        pa.field("id", pa.int64(), nullable=False),
+        pa.field("region", pa.utf8(), nullable=False),
+    ]
+)
+
+
+def _reader(ids):
+    batch = pa.RecordBatch.from_arrays(
+        [
+            pa.array(ids, type=pa.int64()),
+            pa.array(list(range(len(ids))), type=pa.int64()),
+        ],
+        schema=SCHEMA,
+    )
+    return pa.RecordBatchReader.from_batches(SCHEMA, [batch])
+
+
+def _region_reader(rows):
+    batch = pa.RecordBatch.from_arrays(
+        [
+            pa.array([row[0] for row in rows], type=pa.int64()),
+            pa.array([row[1] for row in rows], type=pa.utf8()),
+        ],
+        schema=REGION_SCHEMA,
+    )
+    return pa.RecordBatchReader.from_batches(REGION_SCHEMA, [batch])
+
+
+def _bucket_table(tmp_path):
+    """A table with ``id`` as the primary key and a single-bucket LSM spec."""
+    db = lancedb.connect(tmp_path, read_consistency_interval=timedelta(seconds=0))
+    table = db.create_table("t", _reader([1, 2, 3]))
+    table.set_unenforced_primary_key("id")
+    # num_buckets = 1: every row routes to the single bucket.
+    table.set_lsm_write_spec(LsmWriteSpec.bucket("id", 1))
+    return table
+
+
+def test_lsm_merge_insert_bucket(tmp_path):
+    table = _bucket_table(tmp_path)
+    # Empty `on` defaults to the primary key.
+    result = (
+        table.merge_insert([])
+        .when_matched_update_all()
+        .when_not_matched_insert_all()
+        .execute(_reader([3, 4, 5]))
+    )
+    # LSM path: rows go to the MemWAL, so only num_rows is populated.
+    assert result.num_rows == 3
+    assert result.version == 0
+    assert result.num_inserted_rows == 0
+    assert result.num_updated_rows == 0
+
+
+def test_lsm_merge_insert_unsharded(tmp_path):
+    db = lancedb.connect(tmp_path, read_consistency_interval=timedelta(seconds=0))
+    table = db.create_table("t", _reader([1, 2, 3]))
+    table.set_unenforced_primary_key("id")
+    table.set_lsm_write_spec(LsmWriteSpec.unsharded())
+    result = (
+        table.merge_insert("id")
+        .when_matched_update_all()
+        .when_not_matched_insert_all()
+        .execute(_reader([10, 11, 12, 13]))
+    )
+    assert result.num_rows == 4
+
+
+def test_lsm_merge_insert_identity(tmp_path):
+    db = lancedb.connect(tmp_path, read_consistency_interval=timedelta(seconds=0))
+    table = db.create_table("t", _region_reader([(1, "us"), (2, "us")]))
+    table.set_unenforced_primary_key("id")
+    table.set_lsm_write_spec(LsmWriteSpec.identity("region"))
+    # All rows share one identity value, so they route to one shard.
+    result = (
+        table.merge_insert([])
+        .when_matched_update_all()
+        .when_not_matched_insert_all()
+        .execute(_region_reader([(3, "us"), (4, "us")]))
+    )
+    assert result.num_rows == 2
+
+
+def test_lsm_merge_insert_use_lsm_write_false(tmp_path):
+    table = _bucket_table(tmp_path)  # rows id = 1, 2, 3
+    # use_lsm_write(False) opts out: the standard path runs and commits.
+    result = (
+        table.merge_insert("id")
+        .when_not_matched_insert_all()
+        .use_lsm_write(False)
+        .execute(_reader([3, 4, 5]))
+    )
+    assert result.num_inserted_rows == 2
+    assert table.count_rows() == 5
+
+
+def test_lsm_merge_insert_validate_single_shard_off(tmp_path):
+    table = _bucket_table(tmp_path)
+    result = (
+        table.merge_insert([])
+        .when_matched_update_all()
+        .when_not_matched_insert_all()
+        .validate_single_shard(False)
+        .execute(_reader([6, 7, 8]))
+    )
+    assert result.num_rows == 3
+
+
+def test_lsm_merge_insert_use_lsm_write_true_requires_spec(tmp_path):
+    # A table with a primary key but no LSM write spec installed.
+    db = lancedb.connect(tmp_path, read_consistency_interval=timedelta(seconds=0))
+    table = db.create_table("t", _reader([1, 2, 3]))
+    table.set_unenforced_primary_key("id")
+    with pytest.raises(Exception, match="use_lsm_write"):
+        (
+            table.merge_insert("id")
+            .when_matched_update_all()
+            .when_not_matched_insert_all()
+            .use_lsm_write(True)
+            .execute(_reader([4]))
+        )
+
+
+def test_lsm_merge_insert_rejects_on_not_primary_key(tmp_path):
+    table = _bucket_table(tmp_path)
+    with pytest.raises(Exception, match="primary key"):
+        (
+            table.merge_insert("value")
+            .when_matched_update_all()
+            .when_not_matched_insert_all()
+            .execute(_reader([1]))
+        )
+
+
+def test_lsm_merge_insert_rejects_non_upsert(tmp_path):
+    table = _bucket_table(tmp_path)
+    # Insert-only (no when_matched_update_all) is not the upsert shape.
+    with pytest.raises(Exception, match="upsert"):
+        table.merge_insert([]).when_not_matched_insert_all().execute(_reader([4]))
+
+
+def test_lsm_close_writers(tmp_path):
+    table = _bucket_table(tmp_path)
+    (
+        table.merge_insert([])
+        .when_matched_update_all()
+        .when_not_matched_insert_all()
+        .execute(_reader([7, 8]))
+    )
+    table.close_lsm_writers()
+    # The writer reopens lazily on the next merge_insert.
+    result = (
+        table.merge_insert([])
+        .when_matched_update_all()
+        .when_not_matched_insert_all()
+        .execute(_reader([9]))
+    )
+    assert result.num_rows == 1
+
+
+@pytest.mark.asyncio
+async def test_async_lsm_merge_insert(tmp_path):
+    db = await lancedb.connect_async(
+        tmp_path, read_consistency_interval=timedelta(seconds=0)
+    )
+    table = await db.create_table("t", _reader([1, 2, 3]))
+    await table.set_unenforced_primary_key("id")
+    await table.set_lsm_write_spec(LsmWriteSpec.bucket("id", 1))
+
+    builder = (
+        table.merge_insert([]).when_matched_update_all().when_not_matched_insert_all()
+    )
+    result = await builder.execute(_reader([3, 4, 5]))
+    assert result.num_rows == 3
+    await table.close_lsm_writers()

python/python/tests/test_namespace.py

@@ -5,10 +5,67 @@
 
 import tempfile
 import shutil
+import sys
 import pytest
 import pyarrow as pa
 import lancedb
 from lance_namespace.errors import NamespaceNotEmptyError, TableNotFoundError
+from lancedb.table import AsyncTable, LanceTable
+
+
+PUSHDOWN_DATA = pa.table(
+    {"id": list(range(12)), "text": [f"row-{idx}" for idx in range(12)]}
+)
+
+
+def _ipc_file(table: pa.Table = PUSHDOWN_DATA) -> bytes:
+    sink = pa.BufferOutputStream()
+    with pa.ipc.new_file(sink, table.schema) as writer:
+        writer.write_table(table)
+    return sink.getvalue().to_pybytes()
+
+
+class _FailingSyncInner:
+    name = "hist"
+
+    def current_branch(self):
+        # The pushdown gate only routes server-side when on the default branch.
+        return None
+
+    async def schema(self):
+        return PUSHDOWN_DATA.schema
+
+    async def to_arrow(self):
+        raise RuntimeError("direct table to_arrow should not be used")
+
+
+class _FailingAsyncInner:
+    def name(self):
+        return "hist"
+
+    async def schema(self):
+        return PUSHDOWN_DATA.schema
+
+    def query(self):
+        raise AssertionError("direct async query should not be used")
+
+
+class _NamespaceClient:
+    def __init__(self):
+        self.requests = []
+
+    def query_table(self, request):
+        self.requests.append(request)
+        return _ipc_file()
+
+
+def _namespace_lance_table(namespace_client: _NamespaceClient) -> LanceTable:
+    table = LanceTable.__new__(LanceTable)
+    table._table = _FailingSyncInner()
+    table._namespace_path = ["geneva"]
+    table._namespace_client = namespace_client
+    table._pushdown_operations = {"QueryTable"}
+    return table
 
 
 class TestNamespaceConnection:
@@ -76,6 +133,35 @@ class TestNamespaceConnection:
         assert len(result) == 0
         assert list(result.columns) == ["id", "vector", "text"]
 
+    def test_table_to_pandas_blob_lazy_through_namespace(self):
+        """Namespace-backed tables should use Lance blob-aware pandas conversion."""
+        pytest.importorskip("lance")
+        db = lancedb.connect_namespace("dir", {"root": self.temp_dir})
+        db.create_namespace(["test_ns"])
+        data = pa.table(
+            {
+                "id": pa.array([1, 2], pa.int64()),
+                "blob": pa.array([b"hello", b"world"], pa.large_binary()),
+            },
+            schema=pa.schema(
+                [
+                    pa.field("id", pa.int64()),
+                    pa.field(
+                        "blob",
+                        pa.large_binary(),
+                        metadata={"lance-encoding:blob": "true"},
+                    ),
+                ]
+            ),
+        )
+
+        table = db.create_table("blob_table", data, namespace_path=["test_ns"])
+        df = table.to_pandas(blob_mode="lazy").sort_values("id")
+
+        blob = df["blob"].iloc[0]
+        assert hasattr(blob, "readall")
+        assert blob.readall() == b"hello"
+
     def test_open_table_through_namespace(self):
         """Test opening an existing table through namespace."""
         db = lancedb.connect_namespace("dir", {"root": self.temp_dir})
@@ -707,6 +793,22 @@ class TestPushdownOperations:
         db = lancedb.connect_namespace("dir", {"root": self.temp_dir})
         assert len(db._namespace_client_pushdown_operations) == 0
 
+    def test_lance_table_to_arrow_uses_query_pushdown(self):
+        namespace_client = _NamespaceClient()
+        table = _namespace_lance_table(namespace_client)
+
+        assert table.to_arrow().equals(PUSHDOWN_DATA)
+        assert table.to_pandas()["id"].tolist() == list(range(12))
+        assert len(namespace_client.requests) == 2
+        assert [request.id for request in namespace_client.requests] == [
+            ["geneva", "hist"],
+            ["geneva", "hist"],
+        ]
+        assert [request.k for request in namespace_client.requests] == [
+            sys.maxsize,
+            sys.maxsize,
+        ]
+
 
 @pytest.mark.asyncio
 class TestAsyncPushdownOperations:
@@ -742,3 +844,39 @@ class TestAsyncPushdownOperations:
         """Test that pushdown operations default to empty on async connection."""
         db = lancedb.connect_namespace_async("dir", {"root": self.temp_dir})
         assert len(db._namespace_client_pushdown_operations) == 0
+
+    async def test_async_table_to_arrow_uses_query_pushdown(self):
+        namespace_client = _NamespaceClient()
+
+        table = AsyncTable(
+            _FailingAsyncInner(),
+            namespace_path=["geneva"],
+            namespace_client=namespace_client,
+            pushdown_operations={"QueryTable"},
+        )
+
+        assert (await table.to_arrow()).equals(PUSHDOWN_DATA)
+        assert (await table.to_pandas())["id"].tolist() == list(range(12))
+        assert len(namespace_client.requests) == 2
+        assert [request.id for request in namespace_client.requests] == [
+            ["geneva", "hist"],
+            ["geneva", "hist"],
+        ]
+        assert [request.k for request in namespace_client.requests] == [
+            sys.maxsize,
+            sys.maxsize,
+        ]
+
+
+def test_local_table_to_arrow_and_to_pandas_are_unchanged(tmp_path):
+    db = lancedb.connect(str(tmp_path / "db"))
+    table = db.create_table(
+        "local",
+        data=[
+            {"id": 1, "vector": [1.0, 2.0]},
+            {"id": 2, "vector": [3.0, 4.0]},
+        ],
+    )
+
+    assert table.to_arrow().column("id").to_pylist() == [1, 2]
+    assert table.to_pandas()["id"].tolist() == [1, 2]

python/python/tests/test_query.py

@@ -39,6 +39,35 @@ from utils import exception_output
 from importlib.util import find_spec
 
 
+def _blob_query_data():
+    return pa.table(
+        {
+            "id": pa.array([1, 2, 3, 4], pa.int64()),
+            "tag": pa.array(["drop", "keep", "keep", "keep"], pa.utf8()),
+            "vector": pa.array(
+                [[1.0, 0.0], [2.0, 0.0], [3.0, 0.0], [4.0, 0.0]],
+                type=pa.list_(pa.float32(), list_size=2),
+            ),
+            "blob": pa.array([b"one", b"two", b"three", b"four"], pa.large_binary()),
+        },
+        schema=pa.schema(
+            [
+                pa.field("id", pa.int64()),
+                pa.field("tag", pa.utf8()),
+                pa.field("vector", pa.list_(pa.float32(), list_size=2)),
+                pa.field(
+                    "blob", pa.large_binary(), metadata={"lance-encoding:blob": "true"}
+                ),
+            ]
+        ),
+    )
+
+
+def _assert_lazy_blob(value, expected: bytes):
+    assert hasattr(value, "readall")
+    assert value.readall() == expected
+
+
 @pytest.fixture(scope="module")
 def table(tmpdir_factory) -> lancedb.table.Table:
     tmp_path = str(tmpdir_factory.mktemp("data"))
@@ -181,6 +210,216 @@ async def test_query_to_pandas_kwargs(table, table_async):
     assert async_df["id"].tolist() == [1, 2]
 
 
+@pytest.mark.parametrize("blob_mode", ["lazy", "bytes", "descriptions"])
+def test_plain_scan_query_to_pandas_blob_modes(tmp_db, blob_mode):
+    pytest.importorskip("lance")
+    table = tmp_db.create_table(
+        f"test_query_to_pandas_blob_{blob_mode}", _blob_query_data()
+    )
+
+    df = (
+        table.search()
+        .select(["id", "blob"])
+        .where("id = 1")
+        .to_pandas(blob_mode=blob_mode)
+    )
+
+    assert df["id"].tolist() == [1]
+    if blob_mode == "lazy":
+        _assert_lazy_blob(df["blob"].iloc[0], b"one")
+    elif blob_mode == "bytes":
+        assert df["blob"].tolist() == [b"one"]
+    else:
+        first = df["blob"].iloc[0]
+        assert first != b"one"
+        assert not hasattr(first, "readall")
+
+
+def test_plain_scan_query_to_pandas_blob_projection(tmp_db):
+    pytest.importorskip("lance")
+    table = tmp_db.create_table(
+        "test_query_to_pandas_blob_projection", _blob_query_data()
+    )
+
+    df = (
+        table.search()
+        .where("id >= 2")
+        .select({"id_alias": "id", "payload": "blob", "double_id": "id * 2"})
+        .limit(2)
+        .offset(1)
+        .to_pandas(blob_mode="bytes")
+    )
+
+    assert df["id_alias"].tolist() == [3, 4]
+    assert df["payload"].tolist() == [b"three", b"four"]
+    assert df["double_id"].tolist() == [6, 8]
+
+
+@pytest.mark.parametrize("blob_mode", ["bytes", "descriptions"])
+def test_plain_scan_query_to_pandas_blob_mode_does_not_collect_arrow(
+    tmp_db, monkeypatch, blob_mode
+):
+    pytest.importorskip("lance")
+    table = tmp_db.create_table(
+        "test_query_to_pandas_blob_no_arrow_collect", _blob_query_data()
+    )
+    query = table.search().where("id = 1").select(["id", "blob"])
+
+    def fail_to_arrow(*args, **kwargs):
+        raise AssertionError("to_arrow should not be called before native pandas")
+
+    monkeypatch.setattr(query, "to_arrow", fail_to_arrow)
+
+    df = query.to_pandas(blob_mode=blob_mode)
+
+    assert df["id"].tolist() == [1]
+    if blob_mode == "bytes":
+        assert df["blob"].tolist() == [b"one"]
+    else:
+        first = df["blob"].iloc[0]
+        assert first != b"one"
+        assert not hasattr(first, "readall")
+
+
+def test_plain_scan_query_to_pandas_blob_descriptions_flatten_uses_scanner(
+    tmp_db, monkeypatch
+):
+    pytest.importorskip("lance")
+    table = tmp_db.create_table(
+        "test_query_to_pandas_blob_desc_flatten", _blob_query_data()
+    )
+    query = table.search().where("id = 1").select(["id", "blob"])
+
+    def fail_to_arrow(*args, **kwargs):
+        raise AssertionError("to_arrow should not be called before scanner pandas")
+
+    monkeypatch.setattr(query, "to_arrow", fail_to_arrow)
+
+    df = query.to_pandas(blob_mode="descriptions", flatten=True)
+
+    assert df["id"].tolist() == [1]
+    assert any(column == "blob" or column.startswith("blob.") for column in df.columns)
+
+
+def test_plain_scan_query_to_pandas_scanner_state(tmp_db):
+    pytest.importorskip("lance")
+    data = _blob_query_data()
+    table = tmp_db.create_table("test_query_to_pandas_scanner_state", data.slice(0, 2))
+    table.add(data.slice(2, 2))
+
+    fragments = table.to_lance().get_fragments()
+    assert len(fragments) == 2
+
+    query = (
+        table.search()
+        .select(["id", "blob"])
+        .with_row_address()
+        .fragment_ids([fragments[1].fragment_id])
+    )
+    query_obj = query.to_query_object()
+    assert query_obj.with_row_address is True
+    assert query_obj.fragment_ids == [fragments[1].fragment_id]
+
+    df = query.to_pandas(blob_mode="descriptions")
+
+    assert df["id"].tolist() == [3, 4]
+    assert "_rowaddr" in df.columns
+    assert {rowaddr >> 32 for rowaddr in df["_rowaddr"]} == {fragments[1].fragment_id}
+
+    df_by_fragment = (
+        table.search()
+        .select(["id", "blob"])
+        .with_fragments([fragments[0]])
+        .to_pandas(blob_mode="descriptions")
+    )
+    assert df_by_fragment["id"].tolist() == [1, 2]
+
+
+@pytest.mark.asyncio
+async def test_async_plain_scan_query_to_pandas_blob_projection(tmp_db_async):
+    pytest.importorskip("lance")
+    table = await tmp_db_async.create_table(
+        "test_async_query_to_pandas_blob_projection", _blob_query_data()
+    )
+
+    lazy_df = await (
+        table.query().where("id = 1").select(["id", "blob"]).to_pandas(blob_mode="lazy")
+    )
+    assert lazy_df["id"].tolist() == [1]
+    _assert_lazy_blob(lazy_df["blob"].iloc[0], b"one")
+
+    bytes_df = await (
+        table.query()
+        .where("id >= 2")
+        .select({"id_alias": "id", "payload": "blob", "double_id": "id * 2"})
+        .limit(2)
+        .offset(1)
+        .to_pandas(blob_mode="bytes")
+    )
+    assert bytes_df["id_alias"].tolist() == [3, 4]
+    assert bytes_df["payload"].tolist() == [b"three", b"four"]
+    assert bytes_df["double_id"].tolist() == [6, 8]
+
+    desc_df = await (
+        table.query()
+        .where("id = 1")
+        .select(["blob"])
+        .to_pandas(blob_mode="descriptions")
+    )
+    first = desc_df["blob"].iloc[0]
+    assert first != b"one"
+    assert not hasattr(first, "readall")
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("blob_mode", ["bytes", "descriptions"])
+async def test_async_plain_scan_query_to_pandas_blob_mode_does_not_collect_arrow(
+    tmp_db_async, monkeypatch, blob_mode
+):
+    pytest.importorskip("lance")
+    table = await tmp_db_async.create_table(
+        "test_async_query_to_pandas_blob_no_arrow_collect", _blob_query_data()
+    )
+    query = table.query().where("id = 1").select(["id", "blob"])
+
+    async def fail_to_arrow(*args, **kwargs):
+        raise AssertionError("to_arrow should not be called before native pandas")
+
+    monkeypatch.setattr(query, "to_arrow", fail_to_arrow)
+
+    df = await query.to_pandas(blob_mode=blob_mode)
+
+    assert df["id"].tolist() == [1]
+    if blob_mode == "bytes":
+        assert df["blob"].tolist() == [b"one"]
+    else:
+        first = df["blob"].iloc[0]
+        assert first != b"one"
+        assert not hasattr(first, "readall")
+
+
+def test_vector_query_to_pandas_blob_mode_requires_native_path(tmp_db):
+    pytest.importorskip("lance")
+    table = tmp_db.create_table("test_vector_query_blob_mode", _blob_query_data())
+
+    with pytest.raises(RuntimeError, match="Lance native pandas conversion"):
+        table.search([1.0, 0.0]).select(["blob", "vector"]).limit(1).to_pandas(
+            blob_mode="lazy"
+        )
+
+
+def test_vector_query_to_pandas_blob_descriptions_requires_plain_scan(tmp_db):
+    pytest.importorskip("lance")
+    table = tmp_db.create_table(
+        "test_vector_query_blob_descriptions", _blob_query_data()
+    )
+
+    with pytest.raises(RuntimeError, match="plain scan query"):
+        table.search([1.0, 0.0]).select(["blob", "vector"]).limit(1).to_pandas(
+            blob_mode="descriptions"
+        )
+
+
 def test_order_by_plain_query(mem_db):
     table = mem_db.create_table(
         "test_order_by",

python/python/tests/test_remote_db.py

@@ -1,12 +1,13 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright The LanceDB Authors
-import re
 from concurrent.futures import ThreadPoolExecutor
 import contextlib
 from datetime import timedelta
 import http.server
 import json
 import multiprocessing as mp
+import pickle
+import re
 import sys
 import threading
 import time
@@ -153,6 +154,52 @@ async def test_async_checkout():
         assert await table.count_rows() == 300
 
 
+def test_remote_open_table_branch_and_version():
+    def handler(request):
+        # describe (table open + version validation) always succeeds
+        request.send_response(200)
+        request.send_header("Content-Type", "application/json")
+        request.end_headers()
+        request.wfile.write(
+            json.dumps({"version": 2, "schema": {"fields": []}}).encode()
+        )
+
+    with mock_lancedb_connection(handler) as db:
+        # version-only (and "main" + version) is allowed: remote supports
+        # version time-travel even though it has no branches
+        assert db.open_table("test", version=2) is not None
+        assert db.open_table("test", branch="main", version=2) is not None
+
+        # a non-main branch is rejected, with or without a version
+        with pytest.raises(NotImplementedError, match="branching"):
+            db.open_table("test", branch="exp")
+        with pytest.raises(NotImplementedError, match="branching"):
+            db.open_table("test", branch="exp", version=2)
+
+
+@pytest.mark.asyncio
+async def test_async_remote_open_table_branch_and_version():
+    def handler(request):
+        request.send_response(200)
+        request.send_header("Content-Type", "application/json")
+        request.end_headers()
+        request.wfile.write(
+            json.dumps({"version": 2, "schema": {"fields": []}}).encode()
+        )
+
+    async with mock_lancedb_connection_async(handler) as db:
+        # version-only (and "main" + version) is allowed: "main" is the default
+        # branch, so it must not hit the unsupported remote branch path
+        assert await db.open_table("test", version=2) is not None
+        assert await db.open_table("test", branch="main", version=2) is not None
+
+        # a non-main branch is rejected, with or without a version
+        with pytest.raises(NotImplementedError, match="branching"):
+            await db.open_table("test", branch="exp")
+        with pytest.raises(NotImplementedError, match="branching"):
+            await db.open_table("test", branch="exp", version=2)
+
+
 def test_table_len_sync():
     def handler(request):
         if request.path == "/v1/table/test/create/?mode=create":
@@ -171,6 +218,155 @@ def test_table_len_sync():
         assert len(table) == 1
 
 
+def test_remote_connection_serializes():
+    def handler(request):
+        request.send_response(200)
+        request.send_header("Content-Type", "application/json")
+        request.end_headers()
+        request.wfile.write(b'{"tables": []}')
+
+    with mock_lancedb_connection(handler) as db:
+        serialized = json.loads(db.serialize())
+        assert isinstance(serialized["client_config"], dict)
+        restored = lancedb.deserialize_conn(db.serialize())
+        assert restored.table_names() == []
+
+
+def test_remote_table_is_picklable():
+    def handler(request):
+        request.close_connection = True
+        if request.path == "/v1/table/test/describe/":
+            request.send_response(200)
+            request.send_header("Content-Type", "application/json")
+            request.end_headers()
+            payload = json.dumps(
+                {
+                    "version": 1,
+                    "schema": {
+                        "fields": [
+                            {"name": "id", "type": {"type": "int64"}, "nullable": False}
+                        ]
+                    },
+                }
+            )
+            request.wfile.write(payload.encode())
+        elif request.path == "/v1/table/test/count_rows/":
+            request.send_response(200)
+            request.send_header("Content-Type", "application/json")
+            request.end_headers()
+            request.wfile.write(b"3")
+        else:
+            request.send_response(404)
+            request.end_headers()
+
+    with mock_lancedb_connection(handler) as db:
+        table = db.open_table("test")
+        restored = pickle.loads(pickle.dumps(table))
+        assert restored.count_rows() == 3
+
+
+def test_remote_table_open_does_not_require_picklable_client_config():
+    from lancedb.remote import HeaderProvider
+
+    class LocalHeaderProvider(HeaderProvider):
+        def get_headers(self):
+            return {"X-Test-Header": "present"}
+
+    def handler(request):
+        request.close_connection = True
+        assert request.headers.get("X-Test-Header") == "present"
+        if request.path == "/v1/table/test/describe/":
+            request.send_response(200)
+            request.send_header("Content-Type", "application/json")
+            request.end_headers()
+            request.wfile.write(b'{"version": 1, "schema": {"fields": []}}')
+        elif request.path == "/v1/table/test/count_rows/":
+            request.send_response(200)
+            request.send_header("Content-Type", "application/json")
+            request.end_headers()
+            request.wfile.write(b"3")
+        else:
+            request.send_response(404)
+            request.end_headers()
+
+    with http.server.HTTPServer(
+        ("localhost", 0), make_mock_http_handler(handler)
+    ) as server:
+        port = server.server_address[1]
+        handle = threading.Thread(target=server.serve_forever)
+        handle.start()
+        try:
+            db = lancedb.connect(
+                "db://dev",
+                api_key="fake",
+                host_override=f"http://localhost:{port}",
+                client_config={
+                    "retry_config": {"retries": 0},
+                    "timeout_config": {"connect_timeout": 2, "read_timeout": 2},
+                    "header_provider": LocalHeaderProvider(),
+                },
+            )
+            table = db.open_table("test")
+            assert table.count_rows() == 3
+            with pytest.raises(ValueError, match="header_provider"):
+                pickle.dumps(table)
+        finally:
+            server.shutdown()
+            handle.join()
+
+
+def test_remote_permutation_is_picklable():
+    from lancedb.permutation import Permutation
+
+    rows = list(range(10))
+
+    def handler(request):
+        request.close_connection = True
+        if request.path == "/v1/table/test/describe/":
+            request.send_response(200)
+            request.send_header("Content-Type", "application/json")
+            request.end_headers()
+            payload = json.dumps(
+                {
+                    "version": 1,
+                    "schema": {
+                        "fields": [
+                            {"name": "a", "type": {"type": "int64"}, "nullable": False}
+                        ]
+                    },
+                }
+            )
+            request.wfile.write(payload.encode())
+        elif request.path == "/v1/table/test/count_rows/":
+            request.send_response(200)
+            request.send_header("Content-Type", "application/json")
+            request.end_headers()
+            request.wfile.write(str(len(rows)).encode())
+        elif request.path == "/v1/table/test/query/":
+            content_len = int(request.headers.get("Content-Length"))
+            body = json.loads(request.rfile.read(content_len))
+            if "filter" in body:
+                match = re.search(r"_rowoffset in \((.*?)\)", body["filter"])
+                offsets = [int(offset.strip()) for offset in match.group(1).split(",")]
+            else:
+                offsets = rows
+            table = pa.table({"a": [rows[offset] for offset in offsets]})
+
+            request.send_response(200)
+            request.send_header("Content-Type", "application/vnd.apache.arrow.file")
+            request.end_headers()
+            with pa.ipc.new_file(request.wfile, schema=table.schema) as writer:
+                writer.write_table(table)
+        else:
+            request.send_response(404)
+            request.end_headers()
+
+    with mock_lancedb_connection(handler) as db:
+        permutation = Permutation.identity(db.open_table("test"))
+        restored = pickle.loads(pickle.dumps(permutation))
+        assert restored.__getitems__([0, 2, 4]) == [{"a": 0}, {"a": 2}, {"a": 4}]
+
+
 def test_create_table_exist_ok():
     def handler(request):
         if request.path == "/v1/table/test/create/?mode=exist_ok":
@@ -436,22 +632,25 @@ def test_table_create_indices():
         # This is a smoke-test.
         table = db.create_table("test", [{"id": 1}])
 
-        # Test create_scalar_index with custom name
-        table.create_scalar_index(
-            "id", wait_timeout=timedelta(seconds=2), name="custom_scalar_idx"
-        )
+        # Test create_scalar_index with custom name (legacy method)
+        with pytest.warns(DeprecationWarning, match="create_scalar_index"):
+            table.create_scalar_index(
+                "id", wait_timeout=timedelta(seconds=2), name="custom_scalar_idx"
+            )
 
-        # Test create_fts_index with custom name
-        table.create_fts_index(
-            "text", wait_timeout=timedelta(seconds=2), name="custom_fts_idx"
-        )
+        # Test create_fts_index with custom name (legacy method)
+        with pytest.warns(DeprecationWarning, match="create_fts_index"):
+            table.create_fts_index(
+                "text", wait_timeout=timedelta(seconds=2), name="custom_fts_idx"
+            )
 
-        # Test create_index with custom name
-        table.create_index(
-            vector_column_name="vector",
-            wait_timeout=timedelta(seconds=10),
-            name="custom_vector_idx",
-        )
+        # Test create_index with custom name (legacy form: vector_column_name kwarg)
+        with pytest.warns(DeprecationWarning, match="create_index"):
+            table.create_index(
+                vector_column_name="vector",
+                wait_timeout=timedelta(seconds=10),
+                name="custom_vector_idx",
+            )
 
         # Validate that the name parameter was passed correctly in requests
         assert len(received_requests) == 3
@@ -480,6 +679,98 @@ def test_table_create_indices():
         table.drop_index("custom_fts_idx")
 
 
+def test_remote_create_index_new_api():
+    received_requests = []
+
+    def handler(request):
+        if request.path == "/v1/table/test/create_index/":
+            content_len = int(request.headers.get("Content-Length", 0))
+            body = request.rfile.read(content_len) if content_len > 0 else b""
+            received_requests.append(json.loads(body) if body else {})
+            request.send_response(200)
+            request.end_headers()
+        elif request.path == "/v1/table/test/create/?mode=create":
+            request.send_response(200)
+            request.send_header("Content-Type", "application/json")
+            request.end_headers()
+            request.wfile.write(b"{}")
+        elif request.path == "/v1/table/test/describe/":
+            request.send_response(200)
+            request.send_header("Content-Type", "application/json")
+            request.end_headers()
+            request.wfile.write(
+                json.dumps(
+                    dict(
+                        version=1,
+                        schema=dict(
+                            fields=[
+                                dict(name="id", type={"type": "int64"}, nullable=False),
+                                dict(
+                                    name="category",
+                                    type={"type": "string"},
+                                    nullable=False,
+                                ),
+                                dict(
+                                    name="text", type={"type": "string"}, nullable=False
+                                ),
+                                dict(
+                                    name="vector",
+                                    type={
+                                        "type": "fixed_size_list",
+                                        "fields": [
+                                            dict(
+                                                name="item",
+                                                type={"type": "float"},
+                                                nullable=True,
+                                            )
+                                        ],
+                                        "length": 2,
+                                    },
+                                    nullable=False,
+                                ),
+                            ]
+                        ),
+                    )
+                ).encode()
+            )
+        else:
+            request.send_response(404)
+            request.end_headers()
+
+    from lancedb.index import BTree, FTS, IvfPq, IvfRq
+
+    with mock_lancedb_connection(handler) as db:
+        table = db.create_table("test", [{"id": 1}])
+
+        # New API: column-first, config= kwarg. Should NOT emit DeprecationWarning.
+        import warnings as _warnings
+
+        with _warnings.catch_warnings():
+            _warnings.simplefilter("error", DeprecationWarning)
+            table.create_index("vector", config=IvfPq(distance_type="l2"))
+            table.create_index("category", config=BTree())
+            table.create_index("text", config=FTS())
+            # IvfRq via new API
+            table.create_index("vector", config=IvfRq(distance_type="l2"))
+
+        # Legacy index_type="IVF_RQ" routes to IvfRq config under the hood.
+        with pytest.warns(DeprecationWarning, match="create_index"):
+            table.create_index(
+                vector_column_name="vector",
+                index_type="IVF_RQ",
+                num_partitions=8,
+            )
+
+        assert len(received_requests) == 5
+        assert [req["column"] for req in received_requests] == [
+            "vector",
+            "category",
+            "text",
+            "vector",
+            "vector",
+        ]
+
+
 def test_table_wait_for_index_timeout():
     def handler(request):
         index_stats = dict(
@@ -1305,6 +1596,10 @@ def _remote_fork_child(port: int, queue) -> None:
     queue.put(db.table_names())
 
 
+def _remote_table_fork_child(table, queue) -> None:
+    queue.put(table.count_rows())
+
+
 @pytest.mark.skipif(
     sys.platform != "linux",
     reason=(
@@ -1367,3 +1662,65 @@ def test_remote_connection_after_fork():
     finally:
         server.shutdown()
         server_thread.join()
+
+
+@pytest.mark.skipif(
+    sys.platform != "linux",
+    reason=(
+        "fork() is unavailable on Windows and unsafe on macOS "
+        "(Apple frameworks/TLS are not fork-safe)"
+    ),
+)
+def test_inherited_remote_table_reopens_after_fork():
+    def handler(request):
+        if request.path == "/v1/table/test/describe/":
+            request.send_response(200)
+            request.send_header("Content-Type", "application/json")
+            request.end_headers()
+            request.wfile.write(b'{"version": 1, "schema": {"fields": []}}')
+        elif request.path == "/v1/table/test/count_rows/":
+            request.send_response(200)
+            request.send_header("Content-Type", "application/json")
+            request.end_headers()
+            request.wfile.write(b"7")
+        else:
+            request.send_response(404)
+            request.end_headers()
+
+    server = http.server.HTTPServer(("localhost", 0), make_mock_http_handler(handler))
+    port = server.server_address[1]
+    server_thread = threading.Thread(target=server.serve_forever)
+    server_thread.start()
+    try:
+        db = lancedb.connect(
+            "db://dev",
+            api_key="fake",
+            host_override=f"http://localhost:{port}",
+            client_config={
+                "retry_config": {"retries": 0},
+                "timeout_config": {"connect_timeout": 2, "read_timeout": 2},
+            },
+        )
+        table = db.open_table("test")
+        assert table.count_rows() == 7
+
+        ctx = mp.get_context("fork")
+        queue = ctx.Queue()
+        proc = ctx.Process(target=_remote_table_fork_child, args=(table, queue))
+        proc.start()
+        proc.join(timeout=15)
+
+        if proc.is_alive():
+            proc.terminate()
+            proc.join(timeout=5)
+            if proc.is_alive():
+                proc.kill()
+                proc.join()
+            pytest.fail("Remote table hung after fork")
+
+        assert proc.exitcode == 0, f"child exited with code {proc.exitcode}"
+        assert not queue.empty(), "child produced no result"
+        assert queue.get() == 7
+    finally:
+        server.shutdown()
+        server_thread.join()

python/python/tests/test_rerankers.py

@@ -344,6 +344,12 @@ def test_mrr_reranker(tmp_path):
     assert len(result_deduped) == len(result)
 
 
+def test_mrr_reranker_empty_input():
+    reranker = MRRReranker()
+    with pytest.raises(ValueError, match="must not be empty"):
+        reranker.rerank_multivector([])
+
+
 def test_rrf_reranker_distance():
     data = pa.table(
         {
@@ -603,3 +609,89 @@ def test_cross_encoder_reranker_return_all(tmp_path):
     assert "_relevance_score" in result.column_names
     assert "_score" in result.column_names
     assert "_distance" in result.column_names
+
+
+# ---------------------------------------------------------------------------
+# Regression tests for LinearCombinationReranker scoring bugs (issue #3154)
+# ---------------------------------------------------------------------------
+
+
+def test_linear_combination_best_match_ranks_first():
+    """
+    The document that is BOTH the closest vector match AND the only FTS match
+    must rank first.  Previously _combine_score subtracted from 1, inverting
+    the ranking so the worst document ranked highest.
+    """
+    reranker = LinearCombinationReranker(weight=0.7, return_score="all")
+
+    # rowid 0: perfect vector match, sole FTS match  → should rank 1st
+    # rowid 1: mediocre vector, no FTS match
+    # rowid 2: bad vector, no FTS match
+    vector_results = pa.Table.from_pydict(
+        {
+            "_rowid": [0, 1, 2],
+            "_distance": [0.0, 0.5, 0.9],
+        }
+    )
+    fts_results = pa.Table.from_pydict(
+        {
+            "_rowid": [0],
+            "_score": [1.0],
+        }
+    )
+
+    combined = reranker.merge_results(vector_results, fts_results, fill=1.0)
+    scores = dict(
+        zip(
+            combined["_rowid"].to_pylist(),
+            combined["_relevance_score"].to_pylist(),
+        )
+    )
+
+    # rowid 0 must have the highest relevance score
+    assert scores[0] > scores[1], (
+        f"Best match (rowid 0, score={scores[0]:.4f}) should beat "
+        f"mid match (rowid 1, score={scores[1]:.4f})"
+    )
+    assert scores[1] > scores[2], (
+        f"Mid match (rowid 1, score={scores[1]:.4f}) should beat "
+        f"bad match (rowid 2, score={scores[2]:.4f})"
+    )
+
+
+def test_linear_combination_missing_fts_is_penalised():
+    """
+    A document with no FTS match must score *lower* than a document that
+    has a mediocre FTS match, everything else being equal.  Previously
+    missing-FTS entries used fill=1.0 directly, which gave them a reward
+    (via the 1-(...) inversion) instead of a penalty.
+    """
+    reranker = LinearCombinationReranker(weight=0.5, return_score="all")
+
+    vector_results = pa.Table.from_pydict(
+        {
+            "_rowid": [0, 1],
+            "_distance": [0.2, 0.2],  # identical vector scores
+        }
+    )
+    fts_results = pa.Table.from_pydict(
+        {
+            "_rowid": [0],  # rowid 1 has no FTS match
+            "_score": [0.3],  # small FTS score
+        }
+    )
+
+    combined = reranker.merge_results(vector_results, fts_results, fill=1.0)
+    scores = dict(
+        zip(
+            combined["_rowid"].to_pylist(),
+            combined["_relevance_score"].to_pylist(),
+        )
+    )
+
+    # rowid 0 has a small FTS score; rowid 1 has none.
+    # Even a small FTS contribution should beat having none at all.
+    assert scores[0] > scores[1], (
+        f"Document with FTS score (rowid 0, {scores[0]:.4f}) should beat "
+        f"document with no FTS match (rowid 1, {scores[1]:.4f})"
+    )

python/python/tests/test_table.py

@@ -4,6 +4,8 @@
 
 import os
 import sys
+import threading
+import warnings
 from datetime import date, datetime, timedelta
 from time import sleep
 from typing import List
@@ -11,7 +13,7 @@ from unittest.mock import patch
 
 import lancedb
 from lancedb.dependencies import _PANDAS_AVAILABLE
-from lancedb.index import HnswFlat, HnswPq, HnswSq, IvfPq
+from lancedb.index import BTree, FTS, HnswFlat, HnswPq, HnswSq, IvfPq
 import numpy as np
 import polars as pl
 import pyarrow as pa
@@ -25,6 +27,28 @@ from lancedb.table import LanceTable
 from pydantic import BaseModel
 
 
+def _blob_test_data():
+    return pa.table(
+        {
+            "id": pa.array([1, 2], pa.int64()),
+            "blob": pa.array([b"hello", b"world"], pa.large_binary()),
+        },
+        schema=pa.schema(
+            [
+                pa.field("id", pa.int64()),
+                pa.field(
+                    "blob", pa.large_binary(), metadata={"lance-encoding:blob": "true"}
+                ),
+            ]
+        ),
+    )
+
+
+def _assert_lazy_blob(value, expected: bytes):
+    assert hasattr(value, "readall")
+    assert value.readall() == expected
+
+
 def test_basic(mem_db: DBConnection):
     data = [
         {"vector": [3.1, 4.1], "item": "foo", "price": 10.0},
@@ -56,27 +80,30 @@ def test_table_to_pandas_default_matches_arrow(tmp_db: DBConnection):
     pd.testing.assert_frame_equal(table.to_pandas(), expected)
 
 
-def test_table_to_pandas_blob_bytes(tmp_db: DBConnection):
+def test_table_to_pandas_invalid_blob_mode_non_blob_table(tmp_db: DBConnection):
+    data = pa.table({"id": [1, 2], "text": ["one", "two"]})
+    table = tmp_db.create_table("test_to_pandas_invalid_blob_mode", data=data)
+
+    with pytest.raises(ValueError, match="blob_mode must be one of"):
+        table.to_pandas(blob_mode="invalid")
+
+
+@pytest.mark.parametrize("blob_mode", ["lazy", "bytes", "descriptions"])
+def test_table_to_pandas_blob_modes(tmp_db: DBConnection, blob_mode):
     pytest.importorskip("lance")
-    data = pa.table(
-        {
-            "id": pa.array([1, 2], pa.int64()),
-            "blob": pa.array([b"hello", b"world"], pa.large_binary()),
-        },
-        schema=pa.schema(
-            [
-                pa.field("id", pa.int64()),
-                pa.field(
-                    "blob", pa.large_binary(), metadata={"lance-encoding:blob": "true"}
-                ),
-            ]
-        ),
-    )
-    table = tmp_db.create_table("test_to_pandas_blob_bytes", data=data)
+    table = tmp_db.create_table(f"test_to_pandas_blob_{blob_mode}", _blob_test_data())
 
-    df = table.to_pandas(blob_mode="bytes")
+    df = table.to_pandas(blob_mode=blob_mode)
 
-    assert df["blob"].tolist() == [b"hello", b"world"]
+    if blob_mode == "lazy":
+        _assert_lazy_blob(df["blob"].iloc[0], b"hello")
+        _assert_lazy_blob(df["blob"].iloc[1], b"world")
+    elif blob_mode == "bytes":
+        assert df["blob"].tolist() == [b"hello", b"world"]
+    else:
+        first = df["blob"].iloc[0]
+        assert first != b"hello"
+        assert not hasattr(first, "readall")
 
 
 def test_table_to_pandas_kwargs(tmp_db: DBConnection):
@@ -92,22 +119,8 @@ def test_table_to_pandas_kwargs(tmp_db: DBConnection):
 @pytest.mark.asyncio
 async def test_async_table_to_pandas_blob_bytes(tmp_db_async: AsyncConnection):
     pytest.importorskip("lance")
-    data = pa.table(
-        {
-            "id": pa.array([1, 2], pa.int64()),
-            "blob": pa.array([b"hello", b"world"], pa.large_binary()),
-        },
-        schema=pa.schema(
-            [
-                pa.field("id", pa.int64()),
-                pa.field(
-                    "blob", pa.large_binary(), metadata={"lance-encoding:blob": "true"}
-                ),
-            ]
-        ),
-    )
     table = await tmp_db_async.create_table(
-        "test_async_to_pandas_blob_bytes", data=data
+        "test_async_to_pandas_blob_bytes", data=_blob_test_data()
     )
 
     df = await table.to_pandas(blob_mode="bytes")
@@ -115,6 +128,19 @@ async def test_async_table_to_pandas_blob_bytes(tmp_db_async: AsyncConnection):
     assert df["blob"].tolist() == [b"hello", b"world"]
 
 
+@pytest.mark.asyncio
+async def test_async_table_to_pandas_invalid_blob_mode_non_blob_table(
+    tmp_db_async: AsyncConnection,
+):
+    table = await tmp_db_async.create_table(
+        "test_async_to_pandas_invalid_blob_mode",
+        data=pa.table({"id": [1, 2], "text": ["one", "two"]}),
+    )
+
+    with pytest.raises(ValueError, match="blob_mode must be one of"):
+        await table.to_pandas(blob_mode="invalid")
+
+
 @pytest.mark.asyncio
 async def test_async_table_to_pandas_kwargs(tmp_db_async: AsyncConnection):
     pd = pytest.importorskip("pandas")
@@ -902,6 +928,346 @@ async def test_async_tags(mem_db_async: AsyncConnection):
     )
 
 
+def test_branches(tmp_path):
+    db = lancedb.connect(tmp_path, read_consistency_interval=timedelta(0))
+    table = db.create_table(
+        "test",
+        data=[
+            {"vector": [3.1, 4.1], "item": "foo", "price": 10.0},
+            {"vector": [5.9, 26.5], "item": "bar", "price": 20.0},
+        ],
+    )
+    assert table.count_rows() == 2
+
+    # fork an isolated, writable branch from main
+    branch = table.branches.create("exp")
+    assert branch.count_rows() == 2
+    branch.add(data=[{"vector": [10.0, 11.0], "item": "baz", "price": 30.0}])
+
+    # writes on the branch do not touch main
+    assert branch.count_rows() == 3
+    assert table.count_rows() == 2
+
+    # the branch is listed, with main (None) as its parent
+    branches = table.branches.list()
+    assert "exp" in branches
+    assert branches["exp"]["parent_branch"] is None
+
+    # from_ref="main" is equivalent to the default
+    table.branches.create("exp2", from_ref="main")
+    assert table.branches.list()["exp2"]["parent_branch"] is None
+
+    # checkout returns a handle scoped to the branch's latest
+    checked_out = table.branches.checkout("exp")
+    assert checked_out.count_rows() == 3
+
+    # delete removes it
+    table.branches.delete("exp")
+    table.branches.delete("exp2")
+    assert "exp" not in table.branches.list()
+
+
+def test_branch_handle_tracks_concurrent_writes(tmp_path):
+    db = lancedb.connect(tmp_path, read_consistency_interval=timedelta(0))
+    table = db.create_table("t", [{"id": 1}])
+
+    # two independent handles on the same branch
+    writer = table.branches.create("exp")
+    reader = db.open_table("t", branch="exp")
+    assert reader.count_rows() == 1
+
+    # a concurrent write on the branch is visible to the other handle
+    writer.add([{"id": 2}])
+    assert reader.count_rows() == 2
+    # main is unaffected
+    assert table.count_rows() == 1
+
+
+def test_branch_name_validation(tmp_path):
+    db = lancedb.connect(tmp_path)
+    table = db.create_table("t", [{"id": 1}])
+
+    with pytest.raises(ValueError, match="non-empty"):
+        table.branches.create("")
+    with pytest.raises(ValueError, match="non-empty"):
+        table.branches.checkout("")
+    with pytest.raises(ValueError, match="non-empty"):
+        table.branches.delete("")
+
+
+def test_branches_preserve_namespace(tmp_path):
+    pytest.importorskip(
+        "lance"
+    )  # namespace_path routes through lance's DirectoryNamespace
+    db = lancedb.connect(tmp_path)
+    table = db.create_table("t", [{"id": 1}], namespace_path=["ns1"])
+    assert table.namespace == ["ns1"]
+
+    branch = table.branches.create("exp")
+    assert branch.namespace == ["ns1"]
+    assert branch.id == table.id
+
+    # opening the branch directly also preserves namespace identity
+    opened = db.open_table("t", namespace_path=["ns1"], branch="exp")
+    assert opened.namespace == ["ns1"]
+
+
+def test_open_table_with_branch(tmp_path):
+    db = lancedb.connect(tmp_path)
+    table = db.create_table("t", [{"i": 1}])
+    table.branches.create("exp").add([{"i": 2}])
+
+    # open_table(branch=...) returns a handle scoped to the branch
+    assert db.open_table("t", branch="exp").count_rows() == 2
+    # opening without branch still tracks main
+    assert db.open_table("t").count_rows() == 1
+
+
+def test_open_table_with_branch_version(tmp_path):
+    db = lancedb.connect(tmp_path, read_consistency_interval=timedelta(0))
+
+    # main: a single fork-point row
+    t = db.create_table("t", [{"i": 0}])
+    main_v1 = t.version
+
+    # fork "exp", then advance exp AND main independently past the fork so they
+    # diverge while sharing version numbers
+    exp = t.branches.create("exp")
+    exp.add([{"i": 1}])  # exp: {0, 1}
+    exp_v2 = exp.version
+    exp.add([{"i": 2}])  # exp HEAD: {0, 1, 2}
+    t.add([{"i": 100}, {"i": 101}, {"i": 102}])  # main HEAD: {0, 100, 101, 102}
+    assert exp_v2 == t.version, "branch and main must share the version number"
+
+    # open exp at the shared version: the data must be exp's, not main's. count
+    # alone cannot prove this (main@v2 also exists), so assert provenance by
+    # content.
+    pinned = db.open_table("t", branch="exp", version=exp_v2)
+    assert pinned.current_branch() == "exp"
+    assert pinned.count_rows() == 2  # not exp HEAD (3), not main@v2 (4)
+    assert pinned.count_rows("i = 1") == 1  # exp's post-fork row is visible
+    assert pinned.count_rows("i = 100") == 0  # main's divergent rows are invisible
+
+    # the same coordinate is reachable directly via branches.checkout(name, version)
+    pinned_direct = t.branches.checkout("exp", exp_v2)
+    assert pinned_direct.current_branch() == "exp"
+    assert pinned_direct.count_rows() == 2
+
+    # the HEADs are unaffected
+    assert db.open_table("t", branch="exp").count_rows() == 3
+    assert db.open_table("t").count_rows() == 4
+
+    # version-only (no branch) time-travels main itself: its fork-point version
+    # holds only main's first row, and the shared version number resolves to
+    # main's data, not the branch's ("opens main at the version")
+    old_main = db.open_table("t", version=main_v1)
+    assert old_main.current_branch() is None
+    assert old_main.count_rows() == 1
+    shared_on_main = db.open_table("t", version=exp_v2)
+    assert shared_on_main.current_branch() is None
+    assert shared_on_main.count_rows() == 4
+
+    # detached head: writing to a pinned version is rejected
+    with pytest.raises((ValueError, RuntimeError), match="cannot be modified"):
+        pinned.add([{"i": 9}])
+
+    # a nonexistent version is rejected -- on main, and on a branch (a distinct
+    # resolution path, on the branch's manifests)
+    with pytest.raises((ValueError, RuntimeError)):
+        db.open_table("t", version=9999)
+    with pytest.raises((ValueError, RuntimeError)):
+        db.open_table("t", branch="exp", version=9999)
+
+    # checkout_latest re-attaches the pinned handle to the BRANCH's HEAD
+    # (writable again), not main's HEAD, and not staying pinned
+    pinned.checkout_latest()
+    assert pinned.current_branch() == "exp"
+    assert pinned.count_rows() == 3  # exp HEAD, not main's 4
+    pinned.add([{"i": 3}])
+    assert pinned.count_rows() == 4  # writable again
+
+
+@pytest.mark.asyncio
+async def test_async_namespace_open_table_with_branch(tmp_path):
+    pytest.importorskip("lance")  # "dir" impl is lance.namespace.DirectoryNamespace
+    db = lancedb.connect_namespace_async("dir", {"root": str(tmp_path)})
+    await db.create_namespace(["ns1"])
+    table = await db.create_table("t", [{"id": 1}], namespace_path=["ns1"])
+    branch = await table.branches.create("exp")
+    await branch.add([{"id": 2}])
+
+    # open_table(branch=...) on the async namespace connection must work
+    opened = await db.open_table("t", namespace_path=["ns1"], branch="exp")
+    assert await opened.count_rows() == 2
+
+
+def test_namespace_open_table_with_branch_version(tmp_path):
+    pytest.importorskip("lance")  # "dir" impl is lance.namespace.DirectoryNamespace
+    db = lancedb.connect_namespace("dir", {"root": str(tmp_path)})
+    db.create_namespace(["ns1"])
+    t = db.create_table("t", [{"i": 0}], namespace_path=["ns1"])
+
+    # fork "exp", then advance exp AND main past the fork so they diverge while
+    # sharing version numbers
+    exp = t.branches.create("exp")
+    exp.add([{"i": 1}])
+    exp_v2 = exp.version
+    exp.add([{"i": 2}])
+    t.add([{"i": 100}, {"i": 101}, {"i": 102}])
+    assert exp_v2 == t.version, "branch and main must share the version number"
+
+    # open_table(branch=, version=) on the namespace connection reads the
+    # branch's data at that version, not main's
+    pinned = db.open_table("t", namespace_path=["ns1"], branch="exp", version=exp_v2)
+    assert pinned.current_branch() == "exp"
+    assert pinned.count_rows() == 2  # not exp HEAD (3), not main@v2 (4)
+    assert pinned.count_rows("i = 1") == 1  # exp's post-fork row is visible
+    assert pinned.count_rows("i = 100") == 0  # main's divergent rows are invisible
+    assert db.open_table("t", namespace_path=["ns1"], branch="exp").count_rows() == 3
+
+
+@pytest.mark.asyncio
+async def test_async_namespace_open_table_with_branch_version(tmp_path):
+    pytest.importorskip("lance")  # "dir" impl is lance.namespace.DirectoryNamespace
+    db = lancedb.connect_namespace_async("dir", {"root": str(tmp_path)})
+    await db.create_namespace(["ns1"])
+    t = await db.create_table("t", [{"i": 0}], namespace_path=["ns1"])
+
+    # fork "exp", then advance exp AND main past the fork so they diverge while
+    # sharing version numbers
+    exp = await t.branches.create("exp")
+    await exp.add([{"i": 1}])
+    exp_v2 = await exp.version()
+    await exp.add([{"i": 2}])
+    await t.add([{"i": 100}, {"i": 101}, {"i": 102}])
+    assert exp_v2 == await t.version(), "branch and main must share the version number"
+
+    # open_table(branch=, version=) on the async namespace connection reads the
+    # branch's data at that version, not main's
+    pinned = await db.open_table(
+        "t", namespace_path=["ns1"], branch="exp", version=exp_v2
+    )
+    assert pinned.current_branch() == "exp"
+    assert await pinned.count_rows() == 2  # not exp HEAD (3), not main@v2 (4)
+    assert await pinned.count_rows("i = 1") == 1  # exp's post-fork row is visible
+    assert await pinned.count_rows("i = 100") == 0  # main's rows are invisible
+    assert (
+        await (
+            await db.open_table("t", namespace_path=["ns1"], branch="exp")
+        ).count_rows()
+        == 3
+    )
+
+
+def test_branch_to_lance_targets_branch(tmp_path):
+    pytest.importorskip("lance")
+    db = lancedb.connect(tmp_path)
+    table = db.create_table("t", [{"i": 1}])
+    branch = table.branches.create("exp")
+    branch.add([{"i": 2}])  # branch: 2 rows, main: 1 row
+
+    assert branch.to_lance().count_rows() == 2
+    assert table.to_lance().count_rows() == 1
+
+
+@pytest.mark.asyncio
+async def test_async_branches(tmp_path):
+    db = await lancedb.connect_async(tmp_path)
+    table = await db.create_table(
+        "test",
+        data=[
+            {"vector": [3.1, 4.1], "item": "foo", "price": 10.0},
+            {"vector": [5.9, 26.5], "item": "bar", "price": 20.0},
+        ],
+    )
+    assert await table.count_rows() == 2
+
+    branch = await table.branches.create("exp")
+    assert await branch.count_rows() == 2
+    await branch.add(data=[{"vector": [10.0, 11.0], "item": "baz", "price": 30.0}])
+
+    assert await branch.count_rows() == 3
+    assert await table.count_rows() == 2
+
+    branches = await table.branches.list()
+    assert "exp" in branches
+    assert branches["exp"]["parent_branch"] is None
+
+    await table.branches.create("exp2", from_ref="main")
+    assert (await table.branches.list())["exp2"]["parent_branch"] is None
+
+    checked_out = await table.branches.checkout("exp")
+    assert await checked_out.count_rows() == 3
+
+    await table.branches.delete("exp")
+    await table.branches.delete("exp2")
+    assert "exp" not in await table.branches.list()
+
+
+@pytest.mark.asyncio
+async def test_async_open_table_with_branch_version(tmp_path):
+    db = await lancedb.connect_async(tmp_path, read_consistency_interval=timedelta(0))
+
+    # main: a single fork-point row
+    t = await db.create_table("t", [{"i": 0}])
+    main_v1 = await t.version()
+
+    # fork "exp", then advance exp AND main independently past the fork so they
+    # diverge while sharing version numbers
+    exp = await t.branches.create("exp")
+    await exp.add([{"i": 1}])  # exp: {0, 1}
+    exp_v2 = await exp.version()
+    await exp.add([{"i": 2}])  # exp HEAD: {0, 1, 2}
+    await t.add([{"i": 100}, {"i": 101}, {"i": 102}])  # main HEAD: {0, 100, 101, 102}
+    assert exp_v2 == await t.version(), "branch and main must share the version number"
+
+    # open exp at the shared version: the data must be exp's, not main's. count
+    # alone cannot prove this (main@v2 also exists), so assert provenance by
+    # content.
+    pinned = await db.open_table("t", branch="exp", version=exp_v2)
+    assert pinned.current_branch() == "exp"
+    assert await pinned.count_rows() == 2  # not exp HEAD (3), not main@v2 (4)
+    assert await pinned.count_rows("i = 1") == 1  # exp's post-fork row is visible
+    assert await pinned.count_rows("i = 100") == 0  # main's rows are invisible
+
+    # the same coordinate is reachable directly via branches.checkout(name, version)
+    pinned_direct = await t.branches.checkout("exp", exp_v2)
+    assert pinned_direct.current_branch() == "exp"
+    assert await pinned_direct.count_rows() == 2
+
+    # the HEADs are unaffected
+    assert await (await db.open_table("t", branch="exp")).count_rows() == 3
+    assert await (await db.open_table("t")).count_rows() == 4
+
+    # version-only (no branch) time-travels main itself: its fork-point version
+    # holds only main's first row, and the shared version number resolves to
+    # main's data, not the branch's ("opens main at the version")
+    old_main = await db.open_table("t", version=main_v1)
+    assert old_main.current_branch() is None
+    assert await old_main.count_rows() == 1
+    shared_on_main = await db.open_table("t", version=exp_v2)
+    assert shared_on_main.current_branch() is None
+    assert await shared_on_main.count_rows() == 4
+
+    # detached head: writing to a pinned version is rejected
+    with pytest.raises((ValueError, RuntimeError), match="cannot be modified"):
+        await pinned.add([{"i": 9}])
+
+    # a nonexistent version is rejected -- on main, and on a branch
+    with pytest.raises((ValueError, RuntimeError)):
+        await db.open_table("t", version=9999)
+    with pytest.raises((ValueError, RuntimeError)):
+        await db.open_table("t", branch="exp", version=9999)
+
+    # checkout_latest re-attaches the pinned handle to the BRANCH's HEAD
+    # (writable again), not main's HEAD, and not staying pinned
+    await pinned.checkout_latest()
+    assert pinned.current_branch() == "exp"
+    assert await pinned.count_rows() == 3  # exp HEAD, not main's 4
+    await pinned.add([{"i": 3}])
+    assert await pinned.count_rows() == 4  # writable again
+
+
 @patch("lancedb.table.AsyncTable.create_index")
 def test_create_index_method(mock_create_index, mem_db: DBConnection):
     table = mem_db.create_table(
@@ -928,7 +1294,12 @@ def test_create_index_method(mock_create_index, mem_db: DBConnection):
         num_bits=4,
     )
     mock_create_index.assert_called_with(
-        "vector", replace=True, config=expected_config, name=None, train=True
+        "vector",
+        replace=True,
+        config=expected_config,
+        wait_timeout=None,
+        name=None,
+        train=True,
     )
 
     # Test with target_partition_size
@@ -948,7 +1319,12 @@ def test_create_index_method(mock_create_index, mem_db: DBConnection):
         target_partition_size=8192,
     )
     mock_create_index.assert_called_with(
-        "vector", replace=True, config=expected_config, name=None, train=True
+        "vector",
+        replace=True,
+        config=expected_config,
+        wait_timeout=None,
+        name=None,
+        train=True,
     )
 
     # target_partition_size has a default value,
@@ -967,7 +1343,12 @@ def test_create_index_method(mock_create_index, mem_db: DBConnection):
         num_bits=4,
     )
     mock_create_index.assert_called_with(
-        "vector", replace=True, config=expected_config, name=None, train=True
+        "vector",
+        replace=True,
+        config=expected_config,
+        wait_timeout=None,
+        name=None,
+        train=True,
     )
 
     table.create_index(
@@ -978,7 +1359,12 @@ def test_create_index_method(mock_create_index, mem_db: DBConnection):
     )
     expected_config = HnswPq(distance_type="dot")
     mock_create_index.assert_called_with(
-        "my_vector", replace=False, config=expected_config, name=None, train=True
+        "my_vector",
+        replace=False,
+        config=expected_config,
+        wait_timeout=None,
+        name=None,
+        train=True,
     )
 
     table.create_index(
@@ -993,7 +1379,12 @@ def test_create_index_method(mock_create_index, mem_db: DBConnection):
         distance_type="cosine", sample_rate=0.1, m=29, ef_construction=10
     )
     mock_create_index.assert_called_with(
-        "my_vector", replace=True, config=expected_config, name=None, train=True
+        "my_vector",
+        replace=True,
+        config=expected_config,
+        wait_timeout=None,
+        name=None,
+        train=True,
     )
 
     table.create_index(
@@ -1008,7 +1399,12 @@ def test_create_index_method(mock_create_index, mem_db: DBConnection):
         distance_type="cosine", sample_rate=0.1, m=29, ef_construction=10
     )
     mock_create_index.assert_called_with(
-        "my_vector", replace=True, config=expected_config, name=None, train=True
+        "my_vector",
+        replace=True,
+        config=expected_config,
+        wait_timeout=None,
+        name=None,
+        train=True,
     )
 
 
@@ -1032,6 +1428,7 @@ def test_create_index_name_and_train_parameters(
         "vector",
         replace=True,
         config=expected_config,
+        wait_timeout=None,
         name="my_custom_index",
         train=True,
     )
@@ -1039,13 +1436,82 @@ def test_create_index_name_and_train_parameters(
     # Test with train=False
     table.create_index(vector_column_name="vector", train=False)
     mock_create_index.assert_called_with(
-        "vector", replace=True, config=expected_config, name=None, train=False
+        "vector",
+        replace=True,
+        config=expected_config,
+        wait_timeout=None,
+        name=None,
+        train=False,
     )
 
     # Test with both name and train
     table.create_index(vector_column_name="vector", name="my_index_name", train=True)
     mock_create_index.assert_called_with(
-        "vector", replace=True, config=expected_config, name="my_index_name", train=True
+        "vector",
+        replace=True,
+        config=expected_config,
+        wait_timeout=None,
+        name="my_index_name",
+        train=True,
+    )
+
+
+@patch("lancedb.table.AsyncTable.create_index")
+def test_create_index_legacy_emits_deprecation_warning(
+    mock_create_index, mem_db: DBConnection
+):
+    table = mem_db.create_table(
+        "test",
+        data=[{"vector": [3.1, 4.1]}, {"vector": [5.9, 26.5]}],
+    )
+
+    with pytest.warns(DeprecationWarning, match="create_index"):
+        table.create_index(metric="l2", num_partitions=8, vector_column_name="vector")
+
+
+@patch("lancedb.table.AsyncTable.create_index")
+def test_create_index_new_api(mock_create_index, mem_db: DBConnection):
+    table = mem_db.create_table(
+        "test",
+        data=[
+            {"vector": [3.1, 4.1], "category": "a", "text": "hello world"},
+            {"vector": [5.9, 26.5], "category": "b", "text": "goodbye"},
+        ],
+    )
+
+    # Vector index via new API should not warn
+    with warnings.catch_warnings():
+        warnings.simplefilter("error", DeprecationWarning)
+        table.create_index("vector", config=IvfPq(distance_type="l2"))
+    mock_create_index.assert_called_with(
+        "vector",
+        replace=True,
+        config=IvfPq(distance_type="l2"),
+        wait_timeout=None,
+        name=None,
+        train=True,
+    )
+
+    # Scalar index via new API
+    table.create_index("category", config=BTree())
+    mock_create_index.assert_called_with(
+        "category",
+        replace=True,
+        config=BTree(),
+        wait_timeout=None,
+        name=None,
+        train=True,
+    )
+
+    # FTS index via new API
+    table.create_index("text", config=FTS(with_position=True))
+    mock_create_index.assert_called_with(
+        "text",
+        replace=True,
+        config=FTS(with_position=True),
+        wait_timeout=None,
+        name=None,
+        train=True,
     )
 
 
@@ -1163,6 +1629,45 @@ def test_add_with_empty_fixed_size_list_drops_bad_rows(mem_db: DBConnection):
     assert np.allclose(data["embedding"].to_pylist()[0], np.array([0.1] * 16))
 
 
+def test_add_nullable_struct_with_none(mem_db: DBConnection):
+    """Regression test for issue #2654: a nullable struct column whose
+    first batch contains only None values must not crash in
+    _align_field_types with AttributeError: 'pyarrow.lib.DataType'
+    object has no attribute 'fields'.
+
+    PyArrow infers an all-None struct column as `null` (not `struct`),
+    so the type-alignment path needs to handle the case where the
+    source field type is null and use the target type directly.
+    """
+    # Use the v2.1 file format so that nullable structs are supported.
+    table = mem_db.create_table(
+        "test_nullable_struct",
+        schema=pa.schema(
+            [
+                pa.field("id", pa.string()),
+                pa.field(
+                    "data",
+                    pa.struct([pa.field("x", pa.float32())]),
+                    nullable=True,
+                ),
+            ]
+        ),
+        storage_options=dict(new_table_data_storage_version="2.1"),
+    )
+
+    # Adding a row with a non-null struct should work.
+    table.add([{"id": "1", "data": {"x": 1.0}}])
+
+    # Adding a row with None for the nullable struct field should also
+    # work — this is what used to crash.
+    table.add([{"id": "2", "data": None}])
+
+    result = table.to_arrow()
+    assert result.num_rows == 2
+    assert result.column("id").to_pylist() == ["1", "2"]
+    assert result.column("data").to_pylist() == [{"x": 1.0}, None]
+
+
 def test_add_with_integer_embeddings_preserves_casting(mem_db: DBConnection):
     class Schema(LanceModel):
         text: str
@@ -1861,8 +2366,9 @@ def test_create_scalar_index(mem_db: DBConnection):
         "my_table",
         data=test_data,
     )
-    # Test with default name
-    table.create_scalar_index("x")
+    # Test with default name; confirm DeprecationWarning fires
+    with pytest.warns(DeprecationWarning, match="create_scalar_index"):
+        table.create_scalar_index("x")
     indices = table.list_indices()
     assert len(indices) == 1
     scalar_index = indices[0]
@@ -2370,6 +2876,30 @@ def test_alter_columns(mem_db: DBConnection):
     assert table.to_arrow().column_names == ["new_id"]
 
 
+def test_update_field_metadata(mem_db: DBConnection):
+    data = pa.table({"id": [0, 1], "category": ["a", "b"]})
+    table = mem_db.create_table("my_table", data=data)
+
+    res = table.update_field_metadata(
+        {"path": "category", "metadata": {"unit": "label", "pii": "false"}}
+    )
+    assert res.version == 2
+    # Arrow field metadata is bytes-keyed
+    assert table.schema.field("category").metadata == {
+        b"unit": b"label",
+        b"pii": b"false",
+    }
+
+    # merge: add a key, delete one via None, keep the rest
+    table.update_field_metadata(
+        {"path": "category", "metadata": {"source": "import", "pii": None}}
+    )
+    assert table.schema.field("category").metadata == {
+        b"unit": b"label",
+        b"source": b"import",
+    }
+
+
 @pytest.mark.asyncio
 async def test_alter_columns_async(mem_db_async: AsyncConnection):
     data = pa.table({"id": [0, 1]})
@@ -2648,3 +3178,38 @@ def test_sanitize_data_metadata_not_stripped():
     assert result_schema.metadata is not None
     assert result_schema.metadata[b"existing_key"] == b"existing_value"
     assert result_schema.metadata[b"new_key"] == b"new_value"
+
+
+@pytest.mark.asyncio
+async def test_async_search_runs_embedding_on_dedicated_executor(
+    mem_db_async: AsyncConnection,
+):
+    # Regression test for #3310: AsyncTable.search() must run the (potentially
+    # blocking) query-embedding call on the dedicated embedding executor, not
+    # asyncio's default executor -- which is shared with other blocking I/O and
+    # can be starved by a slow embedding call under concurrent load.
+    func = MockTextEmbeddingFunction.create()
+
+    class Schema(LanceModel):
+        text: str = func.SourceField()
+        vector: Vector(func.ndims()) = func.VectorField()
+
+    table = await mem_db_async.create_table("embed_executor", schema=Schema)
+    await table.add([{"text": "hello world"}])
+
+    captured_threads: List[str] = []
+    original = MockTextEmbeddingFunction.generate_embeddings
+
+    def record_thread(self, texts):
+        captured_threads.append(threading.current_thread().name)
+        return original(self, texts)
+
+    # Patch only around the search so we capture the query-embedding call, not
+    # the add-time source-embedding call.
+    with patch.object(MockTextEmbeddingFunction, "generate_embeddings", record_thread):
+        await (await table.search("a query string")).limit(1).to_list()
+
+    assert captured_threads, "search did not invoke the embedding function"
+    assert all(name.startswith("lancedb-embedding") for name in captured_threads), (
+        f"embedding ran off the dedicated executor: {captured_threads}"
+    )

python/python/tests/test_torch.py

@@ -1,10 +1,15 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright The LanceDB Authors
 
+import contextlib
 import functools
+import http.server
+import json
 import multiprocessing as mp
 import pickle
+import re
 import sys
+import threading
 
 import lancedb
 import pyarrow as pa
@@ -15,6 +20,107 @@ from lancedb.util import tbl_to_tensor
 torch = pytest.importorskip("torch")
 
 
+REMOTE_ROWS = list(range(100))
+
+
+def _make_mock_http_handler(handler):
+    class MockLanceDBHandler(http.server.BaseHTTPRequestHandler):
+        def do_GET(self):
+            handler(self)
+
+        def do_POST(self):
+            handler(self)
+
+    return MockLanceDBHandler
+
+
+def _remote_schema_payload():
+    return {
+        "version": 1,
+        "schema": {
+            "fields": [
+                {"name": "a", "type": {"type": "int64"}, "nullable": False},
+            ]
+        },
+    }
+
+
+def _offsets_from_filter(filter_sql: str | None) -> list[int]:
+    if filter_sql is None:
+        return REMOTE_ROWS
+    match = re.search(r"_rowoffset in \((.*?)\)", filter_sql)
+    if match is None:
+        return REMOTE_ROWS
+    raw_offsets = match.group(1).strip()
+    if raw_offsets == "":
+        return []
+    return [int(offset.strip()) for offset in raw_offsets.split(",")]
+
+
+def _remote_dataset_handler(request):
+    request.close_connection = True
+    if request.path == "/v1/table/test/describe/":
+        request.send_response(200)
+        request.send_header("Content-Type", "application/json")
+        request.end_headers()
+        request.wfile.write(json.dumps(_remote_schema_payload()).encode())
+    elif request.path == "/v1/table/test/count_rows/":
+        request.send_response(200)
+        request.send_header("Content-Type", "application/json")
+        request.end_headers()
+        request.wfile.write(str(len(REMOTE_ROWS)).encode())
+    elif request.path == "/v1/table/test/query/":
+        content_len = int(request.headers.get("Content-Length"))
+        body = json.loads(request.rfile.read(content_len))
+        offsets = _offsets_from_filter(body.get("filter"))
+        requested_columns = body.get("columns") or ["a"]
+        if isinstance(requested_columns, dict):
+            requested_columns = list(requested_columns)
+
+        data = {}
+        for column in requested_columns:
+            if column == "a":
+                data[column] = [REMOTE_ROWS[offset] for offset in offsets]
+            elif column == "_rowoffset":
+                data[column] = offsets
+            elif column == "_rowid":
+                data[column] = offsets
+
+        table = pa.table(data)
+        request.send_response(200)
+        request.send_header("Content-Type", "application/vnd.apache.arrow.file")
+        request.end_headers()
+        with pa.ipc.new_file(request.wfile, schema=table.schema) as writer:
+            writer.write_table(table)
+    else:
+        request.send_response(404)
+        request.end_headers()
+
+
+@contextlib.contextmanager
+def _remote_dataset_table():
+    with http.server.ThreadingHTTPServer(
+        ("localhost", 0), _make_mock_http_handler(_remote_dataset_handler)
+    ) as server:
+        port = server.server_address[1]
+        handle = threading.Thread(target=server.serve_forever)
+        handle.start()
+        try:
+            db = lancedb.connect(
+                "db://dev",
+                api_key="fake",
+                host_override=f"http://localhost:{port}",
+                client_config={
+                    "retry_config": {"retries": 0},
+                    "timeout_config": {"connect_timeout": 2, "read_timeout": 2},
+                },
+            )
+            yield db.open_table("test")
+        finally:
+            server.shutdown()
+            handle.join()
+
+
 def _open_native_table(uri: str, table_name: str):
     """Top-level connection factory used by the explicit-factory pickle test.
 
@@ -107,6 +213,39 @@ def test_permutation_dataloader_multiprocessing(tmp_db):
     assert seen == 1000
 
 
+def test_remote_table_dataloader_multiprocessing():
+    with _remote_dataset_table() as table:
+        dataloader = torch.utils.data.DataLoader(
+            table,
+            collate_fn=tbl_to_tensor,
+            batch_size=10,
+            num_workers=2,
+            multiprocessing_context="spawn",
+        )
+        seen = 0
+        for batch in dataloader:
+            assert batch.size(0) == 1
+            assert batch.size(1) == 10
+            seen += batch.size(1)
+        assert seen == len(REMOTE_ROWS)
+
+
+def test_remote_permutation_dataloader_multiprocessing():
+    with _remote_dataset_table() as table:
+        permutation = Permutation.identity(table)
+        dataloader = torch.utils.data.DataLoader(
+            permutation,
+            batch_size=10,
+            num_workers=2,
+            multiprocessing_context="spawn",
+        )
+        seen = 0
+        for batch in dataloader:
+            assert batch["a"].size(0) == 10
+            seen += batch["a"].size(0)
+        assert seen == len(REMOTE_ROWS)
+
+
 def test_permutation_pickle_with_connection_factory(tmp_path):
     """When the user provides a connection_factory, pickling should round-trip
     through that factory rather than introspecting the connection URI. Useful
@@ -171,6 +310,35 @@ def _multiworker_dataloader_target(db_uri: str, result_queue):
     result_queue.put(count)
 
 
+def _remote_multiworker_dataloader_target(port: int, result_queue):
+    import lancedb
+    from lancedb.permutation import Permutation
+
+    db = lancedb.connect(
+        "db://dev",
+        api_key="fake",
+        host_override=f"http://localhost:{port}",
+        client_config={
+            "retry_config": {"retries": 0},
+            "timeout_config": {"connect_timeout": 2, "read_timeout": 2},
+        },
+    )
+    table = db.open_table("test")
+    permutation = Permutation.identity(table)
+
+    dataloader = torch.utils.data.DataLoader(
+        permutation,
+        batch_size=10,
+        num_workers=2,
+        multiprocessing_context="fork",
+    )
+    count = 0
+    for batch in dataloader:
+        assert batch["a"].size(0) == 10
+        count += 1
+    result_queue.put(count)
+
+
 @pytest.mark.skipif(
     sys.platform != "linux",
     reason=(
@@ -208,3 +376,46 @@ def test_permutation_dataloader_fork_workers(tmp_path):
     assert proc.exitcode == 0, f"child exited with code {proc.exitcode}"
     assert not queue.empty(), "child produced no batches"
     assert queue.get() == 100
+
+
+@pytest.mark.skipif(
+    sys.platform != "linux",
+    reason=(
+        "fork() is unavailable on Windows and unsafe on macOS "
+        "(Apple frameworks/TLS are not fork-safe)"
+    ),
+)
+def test_remote_permutation_dataloader_fork_workers():
+    with http.server.ThreadingHTTPServer(
+        ("localhost", 0), _make_mock_http_handler(_remote_dataset_handler)
+    ) as server:
+        port = server.server_address[1]
+        handle = threading.Thread(target=server.serve_forever)
+        handle.start()
+        try:
+            ctx = mp.get_context("spawn")
+            queue = ctx.Queue()
+            proc = ctx.Process(
+                target=_remote_multiworker_dataloader_target,
+                args=(port, queue),
+            )
+            proc.start()
+            proc.join(timeout=30)
+
+            if proc.is_alive():
+                proc.terminate()
+                proc.join(timeout=5)
+                if proc.is_alive():
+                    proc.kill()
+                    proc.join()
+                pytest.fail(
+                    "Remote permutation hung when iterated in a fork-based "
+                    "DataLoader worker"
+                )
+
+            assert proc.exitcode == 0, f"child exited with code {proc.exitcode}"
+            assert not queue.empty(), "child produced no batches"
+            assert queue.get() == 10
+        finally:
+            server.shutdown()
+            handle.join()

python/python/tests/test_util.py

@@ -149,6 +149,21 @@ def test_value_to_sql_dict():
     assert value_to_sql({}) == "named_struct()"
 
 
+def test_value_to_sql_numpy_scalars():
+    # numpy scalars (e.g. pulled from an ndarray or a pandas column) must
+    # convert the same way as their native Python counterparts. np.float64
+    # already worked by virtue of subclassing float, but the integer / bool
+    # / float32 scalars previously raised NotImplementedError.
+    import numpy as np
+
+    assert value_to_sql(np.int32(5)) == "5"
+    assert value_to_sql(np.int64(5)) == "5"
+    assert value_to_sql(np.float32(1.5)) == "1.5"
+    assert value_to_sql(np.float64(1.5)) == "1.5"
+    assert value_to_sql(np.bool_(True)) == "TRUE"
+    assert value_to_sql(np.bool_(False)) == "FALSE"
+
+
 def test_append_vector_columns():
     registry = EmbeddingFunctionRegistry.get_instance()
     registry.register("test")(MockTextEmbeddingFunction)

python/src/lib.rs

@@ -16,7 +16,7 @@ use query::{FTSQuery, HybridQuery, Query, VectorQuery};
 use session::Session;
 use table::{
     AddColumnsResult, AddResult, AlterColumnsResult, DeleteResult, DropColumnsResult, LsmWriteSpec,
-    MergeResult, Table, UpdateResult,
+    MergeResult, Table, UpdateFieldMetadataResult, UpdateResult,
 };
 
 pub mod arrow;
@@ -50,6 +50,7 @@ pub fn _lancedb(_py: Python, m: &Bound<'_, PyModule>) -> PyResult<()> {
     m.add_class::<RecordBatchStream>()?;
     m.add_class::<AddColumnsResult>()?;
     m.add_class::<AlterColumnsResult>()?;
+    m.add_class::<UpdateFieldMetadataResult>()?;
     m.add_class::<AddResult>()?;
     m.add_class::<MergeResult>()?;
     m.add_class::<LsmWriteSpec>()?;

python/src/table.rs

@@ -16,12 +16,12 @@ use arrow::{
     pyarrow::{FromPyArrow, PyArrowType, ToPyArrow},
 };
 use lancedb::table::{
-    AddDataMode, ColumnAlteration, Duration, NewColumnTransform, OptimizeAction, OptimizeOptions,
-    Table as LanceDbTable,
+    AddDataMode, ColumnAlteration, Duration, FieldMetadataUpdate, NewColumnTransform,
+    OptimizeAction, OptimizeOptions, Ref, Table as LanceDbTable,
 };
 use pyo3::{
     Bound, FromPyObject, Py, PyAny, PyRef, PyResult, Python,
-    exceptions::{PyKeyError, PyRuntimeError, PyValueError},
+    exceptions::{PyRuntimeError, PyValueError},
     pyclass, pymethods,
     types::{IntoPyDict, PyAnyMethods, PyDict, PyDictMethods},
 };
@@ -143,18 +143,20 @@ pub struct MergeResult {
     pub num_inserted_rows: u64,
     pub num_deleted_rows: u64,
     pub num_attempts: u32,
+    pub num_rows: u64,
 }
 
 #[pymethods]
 impl MergeResult {
     pub fn __repr__(&self) -> String {
         format!(
-            "MergeResult(version={}, num_updated_rows={}, num_inserted_rows={}, num_deleted_rows={}, num_attempts={})",
+            "MergeResult(version={}, num_updated_rows={}, num_inserted_rows={}, num_deleted_rows={}, num_attempts={}, num_rows={})",
             self.version,
             self.num_updated_rows,
             self.num_inserted_rows,
             self.num_deleted_rows,
-            self.num_attempts
+            self.num_attempts,
+            self.num_rows
         )
     }
 }
@@ -167,6 +169,7 @@ impl From<lancedb::table::MergeResult> for MergeResult {
             num_inserted_rows: result.num_inserted_rows,
             num_deleted_rows: result.num_deleted_rows,
             num_attempts: result.num_attempts,
+            num_rows: result.num_rows,
         }
     }
 }
@@ -194,6 +197,12 @@ impl LsmWriteSpec {
     }
 
     /// Identity sharding — shard by the raw value of `column`.
+    ///
+    /// `column` must be a deterministic function of the unenforced primary
+    /// key: every row with a given primary key must always produce the same
+    /// `column` value, or upserts of that key can land in different shards
+    /// and a stale version can win. Typically `column` is the primary key
+    /// itself or a stable attribute of it.
     #[staticmethod]
     pub fn identity(column: String) -> Self {
         Self {
@@ -348,6 +357,27 @@ impl From<lancedb::table::AlterColumnsResult> for AlterColumnsResult {
     }
 }
 
+#[pyclass(get_all, from_py_object)]
+#[derive(Clone, Debug)]
+pub struct UpdateFieldMetadataResult {
+    pub version: u64,
+}
+
+#[pymethods]
+impl UpdateFieldMetadataResult {
+    pub fn __repr__(&self) -> String {
+        format!("UpdateFieldMetadataResult(version={})", self.version)
+    }
+}
+
+impl From<lancedb::table::UpdateFieldMetadataResult> for UpdateFieldMetadataResult {
+    fn from(result: lancedb::table::UpdateFieldMetadataResult) -> Self {
+        Self {
+            version: result.version,
+        }
+    }
+}
+
 #[pyclass(get_all, from_py_object)]
 #[derive(Clone, Debug)]
 pub struct DropColumnsResult {
@@ -681,10 +711,6 @@ impl Table {
                         dict.set_item("num_indices", num_indices)?;
                     }
 
-                    if let Some(loss) = stats.loss {
-                        dict.set_item("loss", loss)?;
-                    }
-
                     Ok(Some(dict.unbind()))
                 })
             } else {
@@ -834,6 +860,15 @@ impl Table {
         Ok(Tags::new(self.inner_ref()?.clone()))
     }
 
+    pub fn current_branch(&self) -> PyResult<Option<String>> {
+        Ok(self.inner_ref()?.current_branch())
+    }
+
+    #[getter]
+    pub fn branches(&self) -> PyResult<Branches> {
+        Ok(Branches::new(self.inner_ref()?.clone()))
+    }
+
     #[pyo3(signature = (offsets))]
     pub fn take_offsets(self_: PyRef<'_, Self>, offsets: Vec<u64>) -> PyResult<TakeQuery> {
         Ok(TakeQuery::new(
@@ -933,6 +968,12 @@ impl Table {
         if let Some(use_index) = parameters.use_index {
             builder.use_index(use_index);
         }
+        if let Some(use_lsm_write) = parameters.use_lsm_write {
+            builder.use_lsm_write(use_lsm_write);
+        }
+        if let Some(validate_single_shard) = parameters.validate_single_shard {
+            builder.validate_single_shard(validate_single_shard);
+        }
 
         future_into_py(self_.py(), async move {
             let res = builder.execute(Box::new(batches)).await.infer_error()?;
@@ -971,6 +1012,13 @@ impl Table {
         })
     }
 
+    pub fn close_lsm_writers(self_: PyRef<'_, Self>) -> PyResult<Bound<'_, PyAny>> {
+        let inner = self_.inner_ref()?.clone();
+        future_into_py(self_.py(), async move {
+            inner.close_lsm_writers().await.infer_error()
+        })
+    }
+
     pub fn uses_v2_manifest_paths(self_: PyRef<'_, Self>) -> PyResult<Bound<'_, PyAny>> {
         let inner = self_.inner_ref()?.clone();
         future_into_py(self_.py(), async move {
@@ -1080,31 +1128,57 @@ impl Table {
         field_name: String,
         metadata: &Bound<'_, PyDict>,
     ) -> PyResult<Bound<'a, PyAny>> {
-        let mut new_metadata = HashMap::<String, String>::new();
-        for (column_name, value) in metadata.into_iter() {
-            let key: String = column_name.extract()?;
-            let value: String = value.extract()?;
-            new_metadata.insert(key, value);
+        // Deprecated: forwards to the update_field_metadata path (replace mode).
+        let mut update = FieldMetadataUpdate::new(field_name).replace();
+        for (key, value) in metadata.into_iter() {
+            update = update.set(key.extract::<String>()?, value.extract::<String>()?);
         }
 
         let inner = self_.inner_ref()?.clone();
         future_into_py(self_.py(), async move {
-            let native_tbl = inner
-                .as_native()
-                .ok_or_else(|| PyValueError::new_err("This cannot be run on a remote table"))?;
-            let schema = native_tbl.manifest().await.infer_error()?.schema;
-            let field = schema
-                .field(&field_name)
-                .ok_or_else(|| PyKeyError::new_err(format!("Field {} not found", field_name)))?;
-
-            native_tbl
-                .replace_field_metadata(vec![(field.id as u32, new_metadata)])
-                .await
-                .infer_error()?;
-
+            inner.update_field_metadata(&[update]).await.infer_error()?;
             Ok(())
         })
     }
+
+    pub fn update_field_metadata<'a>(
+        self_: PyRef<'a, Self>,
+        updates: Vec<Bound<PyDict>>,
+    ) -> PyResult<Bound<'a, PyAny>> {
+        let updates = updates
+            .iter()
+            .map(|update| {
+                let path: String = update
+                    .get_item("path")?
+                    .ok_or_else(|| PyValueError::new_err("Missing path"))?
+                    .extract()?;
+                let mut field_update = FieldMetadataUpdate::new(path);
+                if let Some(metadata) = update.get_item("metadata")? {
+                    let metadata_dict = metadata.cast::<PyDict>()?;
+                    for (key, value) in metadata_dict.iter() {
+                        let key: String = key.extract()?;
+                        if value.is_none() {
+                            field_update = field_update.remove(key);
+                        } else {
+                            field_update = field_update.set(key, value.extract::<String>()?);
+                        }
+                    }
+                }
+                if let Some(replace) = update.get_item("replace")?
+                    && replace.extract::<bool>()?
+                {
+                    field_update = field_update.replace();
+                }
+                Ok(field_update)
+            })
+            .collect::<PyResult<Vec<_>>>()?;
+
+        let inner = self_.inner_ref()?.clone();
+        future_into_py(self_.py(), async move {
+            let result = inner.update_field_metadata(&updates).await.infer_error()?;
+            Ok(UpdateFieldMetadataResult::from(result))
+        })
+    }
 }
 
 #[derive(FromPyObject)]
@@ -1124,6 +1198,8 @@ pub struct MergeInsertParams {
     when_not_matched_by_source_condition: Option<String>,
     timeout: Option<std::time::Duration>,
     use_index: Option<bool>,
+    use_lsm_write: Option<bool>,
+    validate_single_shard: Option<bool>,
 }
 
 #[pyclass]
@@ -1194,3 +1270,71 @@ impl Tags {
         })
     }
 }
+
+#[pyclass]
+pub struct Branches {
+    inner: LanceDbTable,
+}
+
+impl Branches {
+    pub fn new(table: LanceDbTable) -> Self {
+        Self { inner: table }
+    }
+}
+
+#[pymethods]
+impl Branches {
+    pub fn list(self_: PyRef<'_, Self>) -> PyResult<Bound<'_, PyAny>> {
+        let inner = self_.inner.clone();
+        future_into_py(self_.py(), async move {
+            let res = inner.list_branches().await.infer_error()?;
+            Python::attach(|py| {
+                let py_dict = PyDict::new(py);
+                for (name, contents) in res {
+                    let value = PyDict::new(py);
+                    value.set_item("parent_branch", contents.parent_branch)?;
+                    value.set_item("parent_version", contents.parent_version)?;
+                    value.set_item("manifest_size", contents.manifest_size)?;
+                    py_dict.set_item(name, value)?;
+                }
+                Ok(py_dict.unbind())
+            })
+        })
+    }
+
+    #[pyo3(signature = (name, from_ref=None, from_version=None))]
+    pub fn create(
+        self_: PyRef<'_, Self>,
+        name: String,
+        from_ref: Option<String>,
+        from_version: Option<u64>,
+    ) -> PyResult<Bound<'_, PyAny>> {
+        let inner = self_.inner.clone();
+        future_into_py(self_.py(), async move {
+            let from = Ref::Version(from_ref, from_version);
+            let table = inner.create_branch(&name, from).await.infer_error()?;
+            Ok(Table::new(table))
+        })
+    }
+
+    #[pyo3(signature = (name, version=None))]
+    pub fn checkout(
+        self_: PyRef<'_, Self>,
+        name: String,
+        version: Option<u64>,
+    ) -> PyResult<Bound<'_, PyAny>> {
+        let inner = self_.inner.clone();
+        future_into_py(self_.py(), async move {
+            let table = inner.checkout_branch(&name, version).await.infer_error()?;
+            Ok(Table::new(table))
+        })
+    }
+
+    pub fn delete(self_: PyRef<'_, Self>, name: String) -> PyResult<Bound<'_, PyAny>> {
+        let inner = self_.inner.clone();
+        future_into_py(self_.py(), async move {
+            inner.delete_branch(&name).await.infer_error()?;
+            Ok(())
+        })
+    }
+}

rust-toolchain.toml

@@ -1,2 +1,2 @@
 [toolchain]
-channel = "1.94.0"
+channel = "1.95.0"

rust/lancedb/Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "lancedb"
-version = "0.30.0-beta.0"
+version = "0.30.1-beta.2"
 edition.workspace = true
 description = "LanceDB: A serverless, low-latency vector database for AI applications"
 license.workspace = true
@@ -75,7 +75,7 @@ reqwest = { version = "0.12.0", default-features = false, features = [
     "stream",
 ], optional = true }
 http = { version = "1", optional = true } # Matching what is in reqwest
-uuid = { version = "1.7.0", features = ["v4"] }
+uuid = { version = "1.7.0", features = ["v4", "v5"] }
 polars-arrow = { version = ">=0.37,<0.40.0", optional = true }
 polars = { version = ">=0.37,<0.40.0", optional = true }
 hf-hub = { version = "0.4.1", optional = true, default-features = false, features = [
@@ -104,6 +104,7 @@ datafusion.workspace = true
 http-body = "1"                                        # Matching reqwest
 rstest = "0.23.0"
 test-log = "0.2"
+serial_test = "3"
 
 
 [features]

rust/lancedb/src/connection.rs

@@ -9,6 +9,7 @@ use std::sync::Arc;
 use arrow_array::RecordBatch;
 use arrow_schema::SchemaRef;
 use lance::dataset::ReadParams;
+use lance::dataset::refs::MAIN_BRANCH;
 use lance_namespace::models::{
     CreateNamespaceRequest, CreateNamespaceResponse, DescribeNamespaceRequest,
     DescribeNamespaceResponse, DropNamespaceRequest, DropNamespaceResponse, ListNamespacesRequest,
@@ -119,6 +120,8 @@ pub struct OpenTableBuilder {
     parent: Arc<dyn Database>,
     request: OpenTableRequest,
     embedding_registry: Arc<dyn EmbeddingRegistry>,
+    branch: Option<String>,
+    version: Option<u64>,
 }
 
 impl OpenTableBuilder {
@@ -139,6 +142,8 @@ impl OpenTableBuilder {
                 managed_versioning: None,
             },
             embedding_registry,
+            branch: None,
+            version: None,
         }
     }
 
@@ -259,14 +264,48 @@ impl OpenTableBuilder {
         self
     }
 
+    /// Open the table scoped to the given branch instead of the default branch.
+    ///
+    /// Reads and writes on the returned table operate in the branch's context.
+    pub fn branch(mut self, branch: impl Into<String>) -> Self {
+        self.branch = Some(branch.into());
+        self
+    }
+
+    /// Open the table pinned to a specific version, producing a read-only "view".
+    ///
+    /// Composes with [`Self::branch`]: when a branch is also set, this opens that
+    /// branch at the given version; otherwise it opens `main` at that version.
+    /// The returned table is a detached head, so operations that modify the table
+    /// will fail until [`Table::checkout_latest`] is called.
+    ///
+    /// ```
+    /// # use lancedb::Connection;
+    /// # async fn f(conn: &Connection) -> Result<(), Box<dyn std::error::Error>> {
+    /// let table = conn.open_table("t").branch("exp").version(3).execute().await?;
+    /// # Ok(())
+    /// # }
+    /// ```
+    pub fn version(mut self, version: u64) -> Self {
+        self.version = Some(version);
+        self
+    }
+
     /// Open the table
     pub async fn execute(self) -> Result<Table> {
         let table = self.parent.open_table(self.request).await?;
-        Ok(Table::new_with_embedding_registry(
-            table,
-            self.parent,
-            self.embedding_registry,
-        ))
+        let table = Table::new_with_embedding_registry(table, self.parent, self.embedding_registry);
+        // "main" is the default branch, so treat it as no branch.
+        let branch = self.branch.filter(|b| b.as_str() != MAIN_BRANCH);
+        match branch {
+            Some(branch) => table.checkout_branch(&branch, self.version).await,
+            None => {
+                if let Some(version) = self.version {
+                    table.checkout(version).await?;
+                }
+                Ok(table)
+            }
+        }
     }
 }
 
@@ -812,8 +851,7 @@ impl ConnectBuilder {
         self
     }
 
-    /// The interval at which to check for updates from other processes. This
-    /// only affects LanceDB OSS.
+    /// The interval at which to check for updates from other processes.
     ///
     /// If left unset, consistency is not checked. For maximum read
     /// performance, this is the default. For strong consistency, set this to
@@ -825,8 +863,11 @@ impl ConnectBuilder {
     /// This only affects read operations. Write operations are always
     /// consistent.
     ///
-    /// LanceDB Cloud uses eventual consistency under the hood, and is not
-    /// currently configurable.
+    /// # Cost
+    ///
+    /// Stronger consistency is not free. The smaller the interval, the more
+    /// often each read pays the cost of checking for updates against object
+    /// storage, raising per-read latency and cost.
     pub fn read_consistency_interval(
         mut self,
         read_consistency_interval: std::time::Duration,
@@ -886,6 +927,7 @@ impl ConnectBuilder {
             options.host_override,
             self.request.client_config,
             storage_options.into(),
+            self.request.read_consistency_interval,
         )?);
         Ok(Connection {
             internal,

rust/lancedb/src/database/namespace.rs

@@ -740,6 +740,64 @@ mod tests {
         assert!(table_names.contains(&"test_table".to_string()));
     }
 
+    #[tokio::test]
+    async fn test_namespace_branch_query_under_pushdown_stays_local() {
+        // With QueryTable pushdown enabled, a query on the main branch routes to
+        // the namespace server, but a branch handle must run locally: the
+        // server-side request carries no branch and would return main's rows.
+        let tmp_dir = tempdir().unwrap();
+        let root_path = tmp_dir.path().to_str().unwrap().to_string();
+
+        let mut properties = HashMap::new();
+        properties.insert("root".to_string(), root_path);
+
+        let conn = connect_namespace("dir", properties)
+            .pushdown_operation(NamespaceClientPushdownOperation::QueryTable)
+            .execute()
+            .await
+            .expect("Failed to connect to namespace");
+
+        conn.create_namespace(CreateNamespaceRequest {
+            id: Some(vec!["test_ns".into()]),
+            ..Default::default()
+        })
+        .await
+        .expect("Failed to create namespace");
+
+        // main has 5 rows
+        let table = conn
+            .create_table("ref_test", create_test_data())
+            .namespace(vec!["test_ns".into()])
+            .execute()
+            .await
+            .expect("Failed to create table");
+        let main_version = table.version().await.unwrap();
+
+        // fork a branch off main, then add 5 more rows so it differs from main
+        let branch = table
+            .create_branch("exp", main_version)
+            .await
+            .expect("Failed to create branch");
+        branch
+            .add(create_test_data())
+            .execute()
+            .await
+            .expect("Failed to append to branch");
+
+        // the branch query must run locally and see the branch's 10 rows --
+        // not get routed to the server (which carries no branch) and see main's 5
+        let results = branch
+            .query()
+            .execute()
+            .await
+            .expect("Failed to query branch")
+            .try_collect::<Vec<_>>()
+            .await
+            .expect("Failed to collect results");
+        let count: usize = results.iter().map(|b| b.num_rows()).sum();
+        assert_eq!(count, 10);
+    }
+
     #[tokio::test]
     async fn test_namespace_describe_table() {
         // Setup: Create a temporary directory for the namespace

rust/lancedb/src/dataloader/permutation/shuffle.rs

@@ -203,11 +203,11 @@ impl Shuffler {
 
         // Finish writing files
         for (file_idx, mut writer) in file_writers.into_iter().enumerate() {
-            let num_written = writer.finish().await?;
+            let write_summary = writer.finish().await?;
             log::debug!(
                 "Shuffle job {}: wrote {} rows to file {}",
                 self.id,
-                num_written,
+                write_summary.num_rows,
                 file_idx
             );
         }
@@ -464,11 +464,9 @@ mod tests {
         let mut iter = ids.into_iter().map(|o| o.unwrap());
         while let Some(first) = iter.next() {
             let rows_left_in_clump = if first == 4470 { 19 } else { 29 };
-            let mut expected_next = first + 1;
-            for _ in 0..rows_left_in_clump {
+            for expected_next in (first + 1)..=(first + rows_left_in_clump) {
                 let next = iter.next().unwrap();
                 assert_eq!(next, expected_next);
-                expected_next += 1;
             }
         }
     }

rust/lancedb/src/index.rs

@@ -372,7 +372,6 @@ pub(crate) struct IndexMetadata {
     pub metric_type: Option<DistanceType>,
     // Sometimes the index type is provided at this level.
     pub index_type: Option<IndexType>,
-    pub loss: Option<f64>,
 }
 
 // This struct is used to deserialize the JSON data returned from the Lance API
@@ -404,6 +403,4 @@ pub struct IndexStatistics {
     pub distance_type: Option<DistanceType>,
     /// The number of parts this index is split into.
     pub num_indices: Option<u32>,
-    /// The loss value used by the index.
-    pub loss: Option<f64>,
 }

rust/lancedb/src/remote/client.rs

@@ -245,6 +245,9 @@ pub struct RestfulLanceDbClient<S: HttpSend = Sender> {
     pub(crate) sender: S,
     pub(crate) id_delimiter: String,
     pub(crate) header_provider: Option<Arc<dyn HeaderProvider>>,
+    /// Connection-level read consistency interval. Drives the
+    /// `x-lancedb-min-timestamp` freshness header sent on read requests.
+    pub(crate) read_consistency_interval: Option<Duration>,
 }
 
 impl<S: HttpSend> std::fmt::Debug for RestfulLanceDbClient<S> {
@@ -338,6 +341,7 @@ impl RestfulLanceDbClient<Sender> {
         host_override: Option<String>,
         default_headers: HeaderMap,
         client_config: ClientConfig,
+        read_consistency_interval: Option<Duration>,
     ) -> Result<Self> {
         // Get the timeouts
         let timeout =
@@ -435,6 +439,7 @@ impl RestfulLanceDbClient<Sender> {
                 .clone()
                 .unwrap_or("$".to_string()),
             header_provider: client_config.header_provider,
+            read_consistency_interval,
         })
     }
 }
@@ -840,6 +845,16 @@ pub mod test_utils {
     pub fn client_with_handler<T>(
         handler: impl Fn(reqwest::Request) -> http::response::Response<T> + Send + Sync + 'static,
     ) -> RestfulLanceDbClient<MockSender>
+    where
+        T: Into<reqwest::Body>,
+    {
+        client_with_handler_and_interval(handler, None)
+    }
+
+    pub fn client_with_handler_and_interval<T>(
+        handler: impl Fn(reqwest::Request) -> http::response::Response<T> + Send + Sync + 'static,
+        read_consistency_interval: Option<Duration>,
+    ) -> RestfulLanceDbClient<MockSender>
     where
         T: Into<reqwest::Body>,
     {
@@ -857,6 +872,7 @@ pub mod test_utils {
             },
             id_delimiter: "$".to_string(),
             header_provider: None,
+            read_consistency_interval,
         }
     }
 
@@ -881,6 +897,7 @@ pub mod test_utils {
             },
             id_delimiter: config.id_delimiter.unwrap_or_else(|| "$".to_string()),
             header_provider: config.header_provider,
+            read_consistency_interval: None,
         }
     }
 }
@@ -888,8 +905,18 @@ pub mod test_utils {
 #[cfg(test)]
 mod tests {
     use super::*;
+    use serial_test::serial;
     use std::time::Duration;
 
+    // Serializes the env-var-mutating tests below: cargo test runs tests in
+    // parallel, but several of these tests read and write the same process-
+    // global env vars (`LANCEDB_USER_ID*`), so they would race without this.
+    static ENV_MUTEX: std::sync::Mutex<()> = std::sync::Mutex::new(());
+
+    fn lock_env() -> std::sync::MutexGuard<'static, ()> {
+        ENV_MUTEX.lock().unwrap_or_else(|e| e.into_inner())
+    }
+
     #[test]
     fn test_timeout_config_default() {
         let config = TimeoutConfig::default();
@@ -1046,6 +1073,7 @@ mod tests {
             sender: Sender,
             id_delimiter: "+".to_string(),
             header_provider: Some(Arc::new(provider) as Arc<dyn HeaderProvider>),
+            read_consistency_interval: None,
         };
 
         // Apply dynamic headers
@@ -1081,6 +1109,7 @@ mod tests {
             sender: Sender,
             id_delimiter: "+".to_string(),
             header_provider: Some(Arc::new(provider) as Arc<dyn HeaderProvider>),
+            read_consistency_interval: None,
         };
 
         // Apply dynamic headers
@@ -1118,6 +1147,7 @@ mod tests {
             sender: Sender,
             id_delimiter: "+".to_string(),
             header_provider: Some(Arc::new(provider) as Arc<dyn HeaderProvider>),
+            read_consistency_interval: None,
         };
 
         // Header provider errors should fail the request
@@ -1143,7 +1173,9 @@ mod tests {
     }
 
     #[test]
+    #[serial(user_id_env)]
     fn test_resolve_user_id_none() {
+        let _guard = lock_env();
         let config = ClientConfig::default();
         // Clear env vars that might be set from other tests
         // SAFETY: This is only called in tests
@@ -1155,7 +1187,9 @@ mod tests {
     }
 
     #[test]
+    #[serial(user_id_env)]
     fn test_resolve_user_id_from_env() {
+        let _guard = lock_env();
         // SAFETY: This is only called in tests
         unsafe {
             std::env::set_var("LANCEDB_USER_ID", "env-user-id");
@@ -1169,7 +1203,9 @@ mod tests {
     }
 
     #[test]
+    #[serial(user_id_env)]
     fn test_resolve_user_id_from_env_key() {
+        let _guard = lock_env();
         // SAFETY: This is only called in tests
         unsafe {
             std::env::remove_var("LANCEDB_USER_ID");
@@ -1189,7 +1225,9 @@ mod tests {
     }
 
     #[test]
+    #[serial(user_id_env)]
     fn test_resolve_user_id_direct_takes_precedence() {
+        let _guard = lock_env();
         // SAFETY: This is only called in tests
         unsafe {
             std::env::set_var("LANCEDB_USER_ID", "env-user-id");
@@ -1206,7 +1244,9 @@ mod tests {
     }
 
     #[test]
+    #[serial(user_id_env)]
     fn test_resolve_user_id_empty_env_ignored() {
+        let _guard = lock_env();
         // SAFETY: This is only called in tests
         unsafe {
             std::env::set_var("LANCEDB_USER_ID", "");

rust/lancedb/src/remote/db.rs

@@ -206,6 +206,7 @@ impl RemoteDatabase {
         host_override: Option<String>,
         client_config: ClientConfig,
         options: RemoteOptions,
+        read_consistency_interval: Option<std::time::Duration>,
     ) -> Result<Self> {
         let parsed = super::client::parse_db_url(uri)?;
         let header_map = RestfulLanceDbClient::<Sender>::default_headers(
@@ -233,6 +234,7 @@ impl RemoteDatabase {
             host_override,
             header_map,
             client_config.clone(),
+            read_consistency_interval,
         )?;
 
         let table_cache = Cache::builder()
@@ -981,6 +983,49 @@ mod tests {
         assert_eq!(table.name(), "table1");
     }
 
+    #[tokio::test]
+    async fn test_open_table_branch_and_version() {
+        // Remote supports version time-travel but not branches. A version-only
+        // open (or one on the default "main" branch) must succeed; a non-main
+        // branch must be rejected, with or without a version.
+        let conn = Connection::new_with_handler(|request| {
+            assert_eq!(request.url().path(), "/v1/table/t/describe/");
+            http::Response::builder()
+                .status(200)
+                .body(
+                    r#"{"table": "t", "version": 2, "schema": {"fields": [
+                        {"name": "a", "type": { "type": "int32" }, "nullable": false}
+                    ]}}"#,
+                )
+                .unwrap()
+        });
+
+        // version-only: allowed (open + checkout(version) both round-trip)
+        conn.open_table("t").version(2).execute().await.unwrap();
+
+        // "main" is the default branch, so it counts as no branch
+        conn.open_table("t")
+            .branch("main")
+            .version(2)
+            .execute()
+            .await
+            .unwrap();
+
+        // a non-main branch is rejected, with or without a version
+        assert!(matches!(
+            conn.open_table("t").branch("exp").execute().await,
+            Err(Error::NotSupported { .. })
+        ));
+        assert!(matches!(
+            conn.open_table("t")
+                .branch("exp")
+                .version(2)
+                .execute()
+                .await,
+            Err(Error::NotSupported { .. })
+        ));
+    }
+
     #[tokio::test]
     async fn test_open_table_not_found() {
         let conn = Connection::new_with_handler(|_| {

rust/lancedb/src/table/add_data.rs

@@ -982,4 +982,105 @@ mod tests {
         table2.add(struct_batch).execute().await.unwrap();
         assert_eq!(table2.count_rows(None).await.unwrap(), 2);
     }
+
+    /// Regression test: appending `arrow.json` (PyArrow `pa.json_()`) data into a table
+    /// whose schema was created with `pa.json_()` (internally stored as `lance.json`, backed
+    /// by `LargeBinary`) must succeed without a schema-mismatch error.
+    ///
+    /// Previously `build_field_exprs` would attempt a `Utf8 → LargeBinary` DataFusion cast,
+    /// which produced a field whose Arrow extension metadata still read `arrow.json` instead
+    /// of `lance.json`.  Lance-core then rejected the append with
+    /// `"json vs large_binary" schema mismatch`.
+    ///
+    /// PyArrow's `pa.json_()` may be backed by either `Utf8` or `LargeUtf8` depending on the
+    /// constructor used, so the test is parameterized over the input backing type.
+    #[rstest::rstest]
+    #[case::utf8(DataType::Utf8)]
+    #[case::large_utf8(DataType::LargeUtf8)]
+    #[tokio::test]
+    async fn test_add_arrow_json_into_lance_json_table(#[case] input_type: DataType) {
+        use arrow_array::{Array, cast::AsArray};
+        use lance_arrow::ARROW_EXT_NAME_KEY;
+        use lance_arrow::json::{ARROW_JSON_EXT_NAME, JSON_EXT_NAME};
+
+        // Build a table whose "data" column is lance.json (LargeBinary +
+        // ARROW:extension:name = "lance.json").
+        let lance_json_field = lance_arrow::json::json_field("data", true);
+        let table_schema = Arc::new(Schema::new(vec![lance_json_field]));
+
+        let db = connect("memory://").execute().await.unwrap();
+        let table = db
+            .create_empty_table("json_test", table_schema)
+            .execute()
+            .await
+            .unwrap();
+
+        // Sanity-check the stored schema.
+        let stored_field = table.schema().await.unwrap();
+        let data_field = stored_field.field_with_name("data").unwrap();
+        assert_eq!(data_field.data_type(), &DataType::LargeBinary);
+        assert_eq!(
+            data_field
+                .metadata()
+                .get(ARROW_EXT_NAME_KEY)
+                .map(|s| s.as_str()),
+            Some(JSON_EXT_NAME),
+        );
+
+        // Build an arrow.json input field (Utf8/LargeUtf8 + arrow.json extension).
+        // This is what PyArrow produces for pa.json_() arrays.
+        let arrow_json_metadata = std::collections::HashMap::from([(
+            ARROW_EXT_NAME_KEY.to_string(),
+            ARROW_JSON_EXT_NAME.to_string(),
+        )]);
+        let arrow_json_field =
+            Field::new("data", input_type.clone(), true).with_metadata(arrow_json_metadata);
+        let arrow_json_schema = Arc::new(Schema::new(vec![arrow_json_field]));
+
+        let rows: Vec<Option<&str>> = vec![None, Some(r#"{"a": 1}"#), Some(r#"{"b": 2}"#)];
+        let string_array: Arc<dyn arrow_array::Array> = match input_type {
+            DataType::Utf8 => Arc::new(arrow_array::StringArray::from(rows.clone())),
+            DataType::LargeUtf8 => Arc::new(arrow_array::LargeStringArray::from(rows.clone())),
+            other => panic!("unsupported arrow.json backing type for this test: {other:?}"),
+        };
+        let batch = RecordBatch::try_new(arrow_json_schema, vec![string_array]).unwrap();
+
+        // This must not fail with a schema-mismatch error.
+        table.add(batch).execute().await.unwrap();
+
+        assert_eq!(table.count_rows(None).await.unwrap(), rows.len());
+
+        // A lance.json column is read back as Utf8 carrying arrow.json extension metadata.
+        let results: Vec<RecordBatch> = table
+            .query()
+            .select(Select::columns(&["data"]))
+            .execute()
+            .await
+            .unwrap()
+            .try_collect()
+            .await
+            .unwrap();
+
+        assert_eq!(results.len(), 1);
+        let batch = &results[0];
+        assert_eq!(batch.num_rows(), rows.len());
+
+        let json_col = batch.column(0);
+        assert_eq!(json_col.data_type(), &DataType::Utf8);
+        let json_strs = json_col.as_string::<i32>();
+
+        for (i, expected) in rows.iter().enumerate() {
+            match expected {
+                None => assert!(json_strs.is_null(i), "row {i} expected null"),
+                Some(raw) => {
+                    assert!(!json_strs.is_null(i), "row {i} expected non-null");
+                    let actual: serde_json::Value = serde_json::from_str(json_strs.value(i))
+                        .expect("read-back JSON should be valid");
+                    let expected: serde_json::Value =
+                        serde_json::from_str(raw).expect("expected JSON should be valid");
+                    assert_eq!(actual, expected, "row {i} JSON mismatch");
+                }
+            }
+        }
+    }
 }

rust/lancedb/src/table/datafusion/cast.rs

@@ -13,6 +13,7 @@ use datafusion_physical_expr::expressions::{CastExpr, Literal};
 use datafusion_physical_plan::expressions::Column;
 use datafusion_physical_plan::projection::ProjectionExec;
 use datafusion_physical_plan::{ExecutionPlan, PhysicalExpr};
+use lance_arrow::json::{is_arrow_json_field, is_json_field};
 
 use crate::{Error, Result};
 
@@ -64,6 +65,18 @@ fn build_field_exprs(
         let input_field = &input_fields[input_idx];
         let input_expr = get_input_expr(input_idx);
 
+        // Special case: input is arrow.json (PyArrow pa.json_() extension type backed by
+        // Utf8/LargeUtf8) and the table field is lance.json (backed by LargeBinary).
+        // Lance-core's write path already handles the arrow.json → lance.json conversion
+        // (including JSONB encoding), so we pass the expression through unchanged and let
+        // lance-core deal with it. Attempting to cast Utf8 → LargeBinary here would
+        // produce a field whose metadata still identifies it as arrow.json, which then
+        // causes a schema-mismatch error inside lance-core.
+        if is_arrow_json_field(input_field) && is_json_field(table_field) {
+            result.push((input_expr, Arc::clone(input_field) as FieldRef));
+            continue;
+        }
+
         let expr = match (input_field.data_type(), table_field.data_type()) {
             // Both are structs: recurse into sub-fields to handle subschemas and casts.
             (DataType::Struct(in_children), DataType::Struct(tbl_children))
@@ -618,4 +631,75 @@ mod tests {
             .unwrap();
         assert_eq!(a.values(), &[1, 3]);
     }
+
+    /// `arrow.json` input (PyArrow `pa.json_()`, Utf8/LargeUtf8 + extension metadata) against a
+    /// `lance.json` table field (LargeBinary + extension metadata) must be passed through
+    /// without a cast so that lance-core can perform its own arrow.json → JSONB conversion.
+    ///
+    /// Before the fix, `cast_to_table_schema` attempted a `Utf8 → LargeBinary` DataFusion
+    /// cast that preserved the wrong extension metadata, causing lance-core to reject the
+    /// batch with a "json vs large_binary" schema-mismatch error.
+    #[rstest::rstest]
+    #[case::utf8(DataType::Utf8)]
+    #[case::large_utf8(DataType::LargeUtf8)]
+    #[tokio::test]
+    async fn test_arrow_json_passthrough_to_lance_json(#[case] input_type: DataType) {
+        use lance_arrow::ARROW_EXT_NAME_KEY;
+        use lance_arrow::json::{ARROW_JSON_EXT_NAME, json_field};
+
+        // Build a table schema with a lance.json field (LargeBinary + lance.json metadata).
+        let lance_field = json_field("data", true);
+        let table_schema = Schema::new(vec![lance_field]);
+
+        // Build an input batch with an arrow.json field (Utf8/LargeUtf8 + arrow.json metadata).
+        let arrow_meta = std::collections::HashMap::from([(
+            ARROW_EXT_NAME_KEY.to_string(),
+            ARROW_JSON_EXT_NAME.to_string(),
+        )]);
+        let arrow_field = Field::new("data", input_type.clone(), true).with_metadata(arrow_meta);
+        let input_schema = Arc::new(Schema::new(vec![arrow_field]));
+
+        let values = vec![Some(r#"{"x": 1}"#), None, Some(r#"{"y": 2}"#)];
+        let input_array: Arc<dyn arrow_array::Array> = match input_type {
+            DataType::Utf8 => Arc::new(StringArray::from(values)),
+            DataType::LargeUtf8 => Arc::new(arrow_array::LargeStringArray::from(values)),
+            other => panic!("unsupported arrow.json backing type for this test: {other:?}"),
+        };
+        let input_batch = RecordBatch::try_new(input_schema, vec![input_array]).unwrap();
+
+        let plan = plan_from_batch(input_batch).await;
+        let projected = cast_to_table_schema(plan, &table_schema).unwrap();
+
+        // The projected schema's "data" field must carry arrow.json metadata
+        // (the input field), not be silently dropped or miscast.
+        let out_field = projected.schema().field_with_name("data").unwrap().clone();
+        assert_eq!(out_field.data_type(), &input_type);
+        assert_eq!(
+            out_field
+                .metadata()
+                .get(ARROW_EXT_NAME_KEY)
+                .map(|s| s.as_str()),
+            Some(ARROW_JSON_EXT_NAME),
+            "output field must still carry arrow.json metadata so lance-core can handle it"
+        );
+
+        // The data must flow through correctly (3 rows, no panic).
+        let result = collect(projected).await;
+        assert_eq!(result.num_rows(), 3);
+        let (v0, v2) = match input_type {
+            DataType::Utf8 => {
+                let col: &StringArray = result.column(0).as_any().downcast_ref().unwrap();
+                (col.value(0).to_string(), col.value(2).to_string())
+            }
+            DataType::LargeUtf8 => {
+                let col: &arrow_array::LargeStringArray =
+                    result.column(0).as_any().downcast_ref().unwrap();
+                (col.value(0).to_string(), col.value(2).to_string())
+            }
+            _ => unreachable!(),
+        };
+        assert_eq!(v0, r#"{"x": 1}"#);
+        assert!(result.column(0).is_null(1));
+        assert_eq!(v2, r#"{"y": 2}"#);
+    }
 }